{ "as_of": "2026-09-30", "count": 2872, "fields": [ "name", "org", "country", "date", "precision", "category", "open_weights", "params", "note", "source" ], "models": [ { "name": "McCulloch-Pitts neuron", "org": "University of Illinois", "country": "USA", "date": "1943-12-01", "precision": "month", "category": "historic", "open_weights": false, "params": "", "note": "First mathematical model of an artificial neuron; foundation of neural networks", "source": "https://en.wikipedia.org/wiki/Timeline_of_machine_learning" }, { "name": "Turochamp", "org": "King's College, Cambridge", "country": "UK", "date": "1948-01-01", "precision": "year", "category": "games", "open_weights": false, "params": "", "note": "Turing & Champernowne chess algorithm; hand-executed against a human in 1952", "source": "https://en.wikipedia.org/wiki/Turochamp" }, { "name": "Machina speculatrix (Elmer and Elsie)", "org": "Burden Neurological Institute", "country": "UK", "date": "1949-01-01", "precision": "year", "category": "robotics", "open_weights": false, "params": "", "note": "Grey Walter's autonomous light-seeking robot tortoises, built 1948-49", "source": "https://en.wikipedia.org/wiki/William_Grey_Walter" }, { "name": "Theseus", "org": "Bell Labs", "country": "USA", "date": "1950-01-01", "precision": "year", "category": "robotics", "open_weights": false, "params": "", "note": "Claude Shannon's maze-learning electromechanical mouse; early machine learning demo", "source": "https://www.technologyreview.com/2018/12/19/138508/mighty-mouse/" }, { "name": "SNARC", "org": "Harvard University", "country": "USA", "date": "1951-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Minsky & Edmonds' first neural network learning machine; documented Jan 1952", "source": "https://en.wikipedia.org/wiki/Stochastic_neural_analog_reinforcement_calculator" }, { "name": "Strachey's draughts program", "org": "University of Manchester", "country": "UK", "date": "1951-01-01", "precision": "year", "category": "games", "open_weights": false, "params": "", "note": "Early checkers program; first ran 1951 on Pilot ACE, then Ferranti Mark 1", "source": "https://en.wikipedia.org/wiki/Christopher_Strachey" }, { "name": "Prinz chess program", "org": "Ferranti", "country": "UK", "date": "1951-11-01", "precision": "month", "category": "games", "open_weights": false, "params": "", "note": "First chess program run on a computer; solved mate-in-two on Ferranti Mark 1", "source": "https://en.wikipedia.org/wiki/Dietrich_Prinz" }, { "name": "Audrey", "org": "Bell Labs", "country": "USA", "date": "1952-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "", "note": "Early speech recognizer: single-speaker spoken digit recognition", "source": "https://en.wikipedia.org/wiki/Speech_recognition" }, { "name": "Georgetown-IBM experiment", "org": "Georgetown University & IBM", "country": "USA", "date": "1954-01-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "First public machine translation demo: 60+ Russian sentences into English", "source": "https://en.wikipedia.org/wiki/Georgetown%E2%80%93IBM_experiment" }, { "name": "Logic Theorist", "org": "RAND Corporation & Carnegie Institute of Technology", "country": "USA", "date": "1956-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Often called the first AI program; proved Principia Mathematica theorems", "source": "https://en.wikipedia.org/wiki/Logic_Theorist" }, { "name": "Samuel Checkers-Playing Program", "org": "IBM", "country": "USA", "date": "1956-02-24", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Self-learning checkers on IBM 701, shown on TV; Samuel coined 'machine learning'", "source": "https://webdocs.cs.ualberta.ca/~chinook/project/legacy.html" }, { "name": "Illiac Suite", "org": "University of Illinois", "country": "USA", "date": "1957-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "", "note": "First musical score composed by an electronic computer (ILLIAC I)", "source": "https://en.wikipedia.org/wiki/Illiac_Suite" }, { "name": "Perceptron", "org": "Cornell Aeronautical Lab", "country": "USA", "date": "1957-01-01", "precision": "month", "category": "historic", "open_weights": false, "params": "", "note": "Rosenblatt's learning neural network, first described in January 1957 report", "source": "https://en.wikipedia.org/wiki/Perceptron" }, { "name": "Pandemonium", "org": "MIT Lincoln Laboratory", "country": "USA", "date": "1958-11-01", "precision": "month", "category": "vision", "open_weights": false, "params": "", "note": "Selfridge's hierarchical 'demons' pattern-recognition architecture, precursor to feature learning", "source": "https://aitopics.org/doc/classics:504E1BAC/" }, { "name": "General Problem Solver", "org": "RAND Corporation & Carnegie Institute of Technology", "country": "USA", "date": "1959-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Newell, Shaw & Simon's universal problem solver using means-ends analysis", "source": "https://en.wikipedia.org/wiki/General_Problem_Solver" }, { "name": "ADALINE", "org": "Stanford University", "country": "USA", "date": "1960-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Widrow-Hoff adaptive linear neuron; introduced the LMS/delta learning rule", "source": "https://isl.stanford.edu/~widrow/papers/c1960adaptiveswitching.pdf" }, { "name": "Mark I Perceptron", "org": "Cornell Aeronautical Lab", "country": "USA", "date": "1960-06-23", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Custom perceptron hardware for image recognition; first public demonstration", "source": "https://en.wikipedia.org/wiki/Perceptron" }, { "name": "'Daisy Bell' singing computer", "org": "Bell Labs", "country": "USA", "date": "1961-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "", "note": "IBM 7090 sang via speech synthesis; earliest computer singing, inspired HAL 9000", "source": "https://en.wikipedia.org/wiki/Daisy_Bell" }, { "name": "MENACE", "org": "University of Edinburgh", "country": "UK", "date": "1961-01-01", "precision": "year", "category": "games", "open_weights": false, "params": "", "note": "Donald Michie's matchbox tic-tac-toe learner; early reinforcement learning", "source": "https://en.wikipedia.org/wiki/Matchbox_Educable_Noughts_and_Crosses_Engine" }, { "name": "SAINT", "org": "MIT", "country": "USA", "date": "1961-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Slagle's symbolic integration program solving college-level calculus problems", "source": "https://en.wikipedia.org/wiki/Timeline_of_artificial_intelligence" }, { "name": "IBM Shoebox", "org": "IBM", "country": "USA", "date": "1962-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "", "note": "16-word speech recognizer debuted at the 1962 Seattle World's Fair", "source": "https://en.wikipedia.org/wiki/Speech_recognition" }, { "name": "STUDENT", "org": "MIT", "country": "USA", "date": "1964-01-01", "precision": "year", "category": "language", "open_weights": false, "params": "", "note": "Bobrow's program solving English-language algebra word problems", "source": "https://en.wikipedia.org/wiki/Timeline_of_artificial_intelligence" }, { "name": "Dendral", "org": "Stanford University", "country": "USA", "date": "1965-01-01", "precision": "year", "category": "science", "open_weights": false, "params": "", "note": "First expert system; inferred molecular structures from mass spectra", "source": "https://en.wikipedia.org/wiki/Timeline_of_artificial_intelligence" }, { "name": "GMDH deep multilayer networks (Ivakhnenko-Lapa)", "org": "Institute of Cybernetics, Kyiv", "country": "Ukraine", "date": "1965-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "First working deep-learning method for multilayer perceptrons (Soviet Ukraine)", "source": "https://en.wikipedia.org/wiki/Timeline_of_artificial_intelligence" }, { "name": "ELIZA", "org": "MIT", "country": "USA", "date": "1966-01-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Weizenbaum's pattern-matching chatbot with DOCTOR script; first famous chatbot", "source": "https://en.wikipedia.org/wiki/ELIZA" }, { "name": "Shakey the robot", "org": "SRI International", "country": "USA", "date": "1966-01-01", "precision": "year", "category": "robotics", "open_weights": false, "params": "", "note": "First mobile robot reasoning about its actions; birthplace of A* and STRIPS", "source": "https://en.wikipedia.org/wiki/Shakey_the_robot" }, { "name": "SYSTRAN", "org": "SYSTRAN", "country": "USA", "date": "1968-01-01", "precision": "year", "category": "language", "open_weights": false, "params": "", "note": "Pioneering rule-based machine translation system, built for US Air Force Russian", "source": "https://en.wikipedia.org/wiki/SYSTRAN" }, { "name": "Reverse-mode automatic differentiation (Linnainmaa)", "org": "University of Helsinki", "country": "Finland", "date": "1970-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Mathematical core of modern backpropagation, published in master's thesis", "source": "https://en.wikipedia.org/wiki/Timeline_of_machine_learning" }, { "name": "SHRDLU", "org": "MIT", "country": "USA", "date": "1970-01-01", "precision": "year", "category": "language", "open_weights": false, "params": "", "note": "Winograd's natural-language blocks-world system; developed 1968-1970", "source": "https://en.wikipedia.org/wiki/SHRDLU" }, { "name": "AARON", "org": "UC San Diego", "country": "USA", "date": "1972-01-01", "precision": "year", "category": "image", "open_weights": false, "params": "", "note": "Harold Cohen's long-running autonomous drawing and painting program", "source": "https://en.wikipedia.org/wiki/AARON" }, { "name": "PARRY", "org": "Stanford University", "country": "USA", "date": "1972-01-01", "precision": "year", "category": "language", "open_weights": false, "params": "", "note": "Colby's paranoid-patient chatbot; conversed with ELIZA over ARPANET in 1972", "source": "https://en.wikipedia.org/wiki/PARRY" }, { "name": "MYCIN", "org": "Stanford University", "country": "USA", "date": "1973-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Rule-based expert system recommending antibiotic therapy for infections", "source": "https://pubmed.ncbi.nlm.nih.gov/4589706/" }, { "name": "WABOT-1", "org": "Waseda University", "country": "Japan", "date": "1973-01-01", "precision": "year", "category": "robotics", "open_weights": false, "params": "", "note": "First full-scale anthropomorphic robot: walked, gripped objects, spoke Japanese", "source": "http://www.humanoid.waseda.ac.jp/booklet/kato_2.html" }, { "name": "Harpy", "org": "Carnegie Mellon University", "country": "USA", "date": "1976-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "", "note": "DARPA-funded speech system understanding 1,011 words", "source": "https://en.wikipedia.org/wiki/Timeline_of_speech_and_voice_recognition" }, { "name": "Kurzweil Reading Machine", "org": "Kurzweil Computer Products", "country": "USA", "date": "1976-01-13", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "First omni-font OCR reading machine, reading printed text aloud for blind users", "source": "https://en.wikipedia.org/wiki/Kurzweil_Computer_Products" }, { "name": "Stanford Cart (Moravec obstacle run)", "org": "Stanford University", "country": "USA", "date": "1979-01-01", "precision": "year", "category": "robotics", "open_weights": false, "params": "", "note": "Autonomously crossed a chair-filled room using stereo vision", "source": "https://en.wikipedia.org/wiki/Timeline_of_artificial_intelligence" }, { "name": "BKG 9.8", "org": "Carnegie Mellon University", "country": "USA", "date": "1979-07-01", "precision": "month", "category": "games", "open_weights": false, "params": "", "note": "First program to beat a reigning world champion (backgammon, Luigi Villa)", "source": "https://en.wikipedia.org/wiki/BKG_9.8" }, { "name": "Neocognitron", "org": "NHK Science & Technical Research Laboratories", "country": "Japan", "date": "1979-10-01", "precision": "month", "category": "vision", "open_weights": false, "params": "", "note": "Fukushima's hierarchical shift-invariant network; direct ancestor of CNNs", "source": "https://en.wikipedia.org/wiki/Neocognitron" }, { "name": "XCON (R1)", "org": "Digital Equipment Corporation & Carnegie Mellon University", "country": "USA", "date": "1980-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "First commercially successful expert system, configuring DEC VAX orders", "source": "https://en.wikipedia.org/wiki/Xcon" }, { "name": "Experiments in Musical Intelligence (EMI)", "org": "UC Santa Cruz", "country": "USA", "date": "1981-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "", "note": "David Cope's program composing music in the style of classical masters", "source": "https://en.wikipedia.org/wiki/David_Cope" }, { "name": "Self-Organizing Map (Kohonen network)", "org": "Helsinki University of Technology", "country": "Finland", "date": "1982-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Unsupervised topology-preserving neural feature maps", "source": "https://link.springer.com/article/10.1007/BF00337288" }, { "name": "Hopfield network", "org": "Caltech", "country": "USA", "date": "1982-04-01", "precision": "month", "category": "historic", "open_weights": false, "params": "", "note": "Recurrent associative-memory network; basis of the 2024 Physics Nobel", "source": "https://www.pnas.org/doi/10.1073/pnas.79.8.2554" }, { "name": "DECtalk", "org": "Digital Equipment Corporation", "country": "USA", "date": "1983-12-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Landmark text-to-speech synthesizer from Klatt's work; announced Dec 1983", "source": "https://en.wikipedia.org/wiki/DECtalk" }, { "name": "Cyc", "org": "MCC (Microelectronics and Computer Technology Corp.)", "country": "USA", "date": "1984-07-01", "precision": "month", "category": "historic", "open_weights": false, "params": "", "note": "Lenat's decades-long project to hand-encode common-sense knowledge", "source": "https://en.wikipedia.org/wiki/Cyc" }, { "name": "Boltzmann machine", "org": "Carnegie Mellon University & Johns Hopkins University", "country": "USA", "date": "1985-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Ackley, Hinton & Sejnowski's stochastic generative network and learning algorithm", "source": "https://en.wikipedia.org/wiki/Boltzmann_machine" }, { "name": "NETtalk", "org": "Johns Hopkins University & Princeton University", "country": "USA", "date": "1986-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "18.6K", "note": "Neural network that learned to pronounce English text aloud", "source": "https://en.wikipedia.org/wiki/NETtalk_(artificial_neural_network)" }, { "name": "VaMoRs", "org": "Bundeswehr University Munich", "country": "Germany", "date": "1986-01-01", "precision": "year", "category": "robotics", "open_weights": false, "params": "", "note": "Dickmanns' vision-guided robot van; drove itself at 96 km/h by 1987", "source": "https://en.wikipedia.org/wiki/Ernst_Dickmanns" }, { "name": "Backpropagation (Rumelhart, Hinton & Williams)", "org": "UC San Diego & Carnegie Mellon University", "country": "USA", "date": "1986-10-09", "precision": "day", "category": "historic", "open_weights": false, "params": "", "note": "Nature paper popularizing backprop for learning internal representations", "source": "https://www.semanticscholar.org/paper/Learning-representations-by-back-propagating-errors-Rumelhart-Hinton/052b1d8ce63b07fec3de9dbb583772d860b7c769" }, { "name": "Time delay neural network (TDNN)", "org": "ATR (Advanced Telecommunications Research Institute)", "country": "Japan", "date": "1987-12-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Waibel's shift-invariant network for phoneme recognition; early convolutional speech model", "source": "https://en.wikipedia.org/wiki/Time_delay_neural_network" }, { "name": "IBM Models 1-5 (statistical machine translation)", "org": "IBM", "country": "USA", "date": "1988-01-01", "precision": "year", "category": "language", "open_weights": false, "params": "", "note": "Founded statistical MT; published 1988-1993, word alignment used for two decades", "source": "https://en.wikipedia.org/wiki/IBM_alignment_models" }, { "name": "ALVINN", "org": "Carnegie Mellon University", "country": "USA", "date": "1988-11-01", "precision": "month", "category": "robotics", "open_weights": false, "params": "", "note": "Neural network steering the Navlab autonomous vehicle from camera input", "source": "https://proceedings.neurips.cc/paper/1988/hash/812b4ba287f5ee0bc9d43bbf5bbe87fb-Abstract.html" }, { "name": "Deep Thought", "org": "Carnegie Mellon University", "country": "USA", "date": "1988-11-01", "precision": "month", "category": "games", "open_weights": false, "params": "", "note": "First computer to beat a grandmaster in tournament play (Bent Larsen)", "source": "https://en.wikipedia.org/wiki/Deep_Thought_(chess_computer)" }, { "name": "Q-learning", "org": "University of Cambridge", "country": "UK", "date": "1989-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Watkins' model-free reinforcement learning algorithm; later basis of DQN", "source": "http://www.cs.rhul.ac.uk/~chrisw/thesis.html" }, { "name": "LeNet (zip code CNN)", "org": "AT&T Bell Labs", "country": "USA", "date": "1989-12-01", "precision": "month", "category": "vision", "open_weights": false, "params": "9.8K", "note": "First backprop-trained convolutional network, reading handwritten ZIP codes", "source": "https://en.wikipedia.org/wiki/LeNet" }, { "name": "Chinook", "org": "University of Alberta", "country": "Canada", "date": "1990-01-01", "precision": "year", "category": "games", "open_weights": false, "params": "", "note": "Checkers program; earned title shot 1990, Man-Machine World Champion 1994", "source": "https://en.wikipedia.org/wiki/Chinook_(computer_program)" }, { "name": "DragonDictate", "org": "Dragon Systems", "country": "USA", "date": "1990-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "", "note": "First consumer speech recognition product (discrete words, HMM-based)", "source": "https://en.wikipedia.org/wiki/Timeline_of_speech_and_voice_recognition" }, { "name": "Elman network (simple recurrent network)", "org": "UC San Diego", "country": "USA", "date": "1990-01-01", "precision": "year", "category": "historic", "open_weights": false, "params": "", "note": "Recurrent network with context units learning temporal structure in sequences", "source": "https://en.wikipedia.org/wiki/Recurrent_neural_network" }, { "name": "Sphinx-II", "org": "Carnegie Mellon University", "country": "USA", "date": "1992-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "", "note": "Speaker-independent large-vocabulary continuous recognizer; won DARPA's 1992 evaluation", "source": "https://en.wikipedia.org/wiki/Speech_recognition" }, { "name": "TD-Gammon", "org": "IBM", "country": "USA", "date": "1992-05-01", "precision": "month", "category": "games", "open_weights": false, "params": "25K", "note": "Temporal-difference neural net reaching near-world-class backgammon via self-play", "source": "https://papers.nips.cc/paper/1991/file/68ce199ec2c5517597ce0a4d89620f55-Paper.pdf" }, { "name": "Navlab 5 / RALPH (No Hands Across America)", "org": "Carnegie Mellon University", "country": "USA", "date": "1995-07-01", "precision": "month", "category": "robotics", "open_weights": false, "params": "", "note": "Vision-steered minivan drove ~98% of 2,850 miles coast to coast", "source": "https://en.wikipedia.org/wiki/Navlab" }, { "name": "Support-vector machine (Cortes & Vapnik)", "org": "AT&T Bell Labs", "country": "USA", "date": "1995-09-01", "precision": "month", "category": "historic", "open_weights": false, "params": "", "note": "Soft-margin support-vector networks; dominant ML method of the 2000s", "source": "https://link.springer.com/article/10.1007/BF00994018" }, { "name": "A.L.I.C.E.", "org": "Richard Wallace", "country": "USA", "date": "1995-11-23", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "AIML chatbot; three-time Loebner Prize winner (2000, 2001, 2004)", "source": "https://en.wikipedia.org/wiki/Artificial_Linguistic_Internet_Computer_Entity" }, { "name": "Deep Blue", "org": "IBM", "country": "USA", "date": "1996-02-10", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "First computer to win a game against a reigning world chess champion", "source": "https://en.wikipedia.org/wiki/Deep_Blue_versus_Garry_Kasparov" }, { "name": "Logistello", "org": "NEC Research Institute", "country": "USA", "date": "1997-01-01", "precision": "year", "category": "games", "open_weights": false, "params": "", "note": "Michael Buro's Othello program; beat world champion Takeshi Murakami 6-0", "source": "https://en.wikipedia.org/wiki/Logistello" }, { "name": "Deep Blue (1997 upgrade)", "org": "IBM", "country": "USA", "date": "1997-05-03", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Upgraded Deep Blue beat Kasparov 3½-2½ in May 3-11 rematch", "source": "https://en.wikipedia.org/wiki/Deep_Blue_versus_Garry_Kasparov" }, { "name": "Dragon NaturallySpeaking 1.0", "org": "Dragon Systems", "country": "USA", "date": "1997-06-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "First consumer continuous-dictation speech recognition software", "source": "https://en.wikipedia.org/wiki/Dragon_NaturallySpeaking" }, { "name": "Bidirectional RNN", "org": "ATR (Advanced Telecommunications Research Institute)", "country": "Japan", "date": "1997-11-01", "precision": "month", "category": "historic", "open_weights": false, "params": "", "note": "Schuster & Paliwal's RNN reading sequences in both directions", "source": "https://ieeexplore.ieee.org/document/650093" }, { "name": "LSTM", "org": "TU Munich & IDSIA", "country": "Germany", "date": "1997-11-15", "precision": "day", "category": "historic", "open_weights": false, "params": "", "note": "Hochreiter & Schmidhuber's gated recurrent network solving vanishing gradients", "source": "https://direct.mit.edu/neco/article-abstract/9/8/1735/6109/Long-Short-Term-Memory" }, { "name": "LeNet-5", "org": "AT&T Labs", "country": "USA", "date": "1998-11-01", "precision": "month", "category": "vision", "open_weights": false, "params": "60K", "note": "Canonical CNN for document recognition; read ~10% of US checks", "source": "https://en.wikipedia.org/wiki/LeNet" }, { "name": "AIBO (ERS-110)", "org": "Sony", "country": "Japan", "date": "1999-05-11", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "First mass-market autonomous robot pet; sold out online in 20 minutes", "source": "https://en.wikipedia.org/wiki/AIBO" }, { "name": "ASIMO", "org": "Honda", "country": "Japan", "date": "2000-10-01", "precision": "month", "category": "robotics", "open_weights": false, "params": "", "note": "Honda's advanced bipedal humanoid robot, unveiled October 2000", "source": "https://en.wikipedia.org/wiki/ASIMO" }, { "name": "Neural Probabilistic Language Model", "org": "Université de Montréal", "country": "Canada", "date": "2000-11-01", "precision": "month", "category": "language", "open_weights": false, "params": "6.9M", "note": "Bengio et al.'s first neural language model with learned word embeddings", "source": "https://papers.nips.cc/paper_files/paper/2000/file/728f206c2a01bf572b5940d7d9a8fa4c-Paper.pdf" }, { "name": "Eugene Goostman", "org": "Vladimir Veselov, Eugene Demchenko & Sergey Ulasen", "country": "Russia", "date": "2001-01-01", "precision": "year", "category": "language", "open_weights": false, "params": "", "note": "Chatbot posing as a 13-year-old; claimed Turing test pass June 2014", "source": "https://en.wikipedia.org/wiki/Eugene_Goostman" }, { "name": "Viola-Jones face detector", "org": "MERL & Compaq CRL", "country": "USA", "date": "2001-12-01", "precision": "month", "category": "vision", "open_weights": false, "params": "", "note": "Boosted cascade enabling real-time face detection; ubiquitous in cameras", "source": "https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/viola-cvpr-01.pdf" }, { "name": "Roomba", "org": "iRobot", "country": "USA", "date": "2002-09-01", "precision": "month", "category": "robotics", "open_weights": false, "params": "", "note": "First mass-market autonomous home robot vacuum", "source": "https://en.wikipedia.org/wiki/Roomba" }, { "name": "Vocaloid", "org": "Yamaha", "country": "Japan", "date": "2003-03-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Singing voice synthesizer announced at Musikmesse 2003; first products January 2004", "source": "https://en.wikipedia.org/wiki/Vocaloid" }, { "name": "BigDog", "org": "Boston Dynamics", "country": "USA", "date": "2005-01-01", "precision": "year", "category": "robotics", "open_weights": false, "params": "", "note": "Dynamically balancing quadruped robot for rough terrain", "source": "https://en.wikipedia.org/wiki/BigDog" }, { "name": "Stanley", "org": "Stanford University", "country": "USA", "date": "2005-10-08", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "ML-driven autonomous VW Touareg; won DARPA Grand Challenge desert race", "source": "https://en.wikipedia.org/wiki/DARPA_Grand_Challenge" }, { "name": "Google Translate (statistical MT)", "org": "Google", "country": "USA", "date": "2006-04-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Launched with Google's own statistical MT; went neural in 2016", "source": "https://en.wikipedia.org/wiki/Google_Translate" }, { "name": "Crazy Stone", "org": "INRIA", "country": "France", "date": "2006-05-01", "precision": "month", "category": "games", "open_weights": false, "params": "", "note": "Rémi Coulom's Go program introducing Monte Carlo tree search; precursor to AlphaGo", "source": "https://inria.hal.science/inria-00116992v1" }, { "name": "Connectionist Temporal Classification (CTC)", "org": "IDSIA", "country": "Switzerland", "date": "2006-06-01", "precision": "month", "category": "historic", "open_weights": false, "params": "", "note": "Graves et al.'s loss for training RNNs on unsegmented speech/handwriting", "source": "https://en.wikipedia.org/wiki/Connectionist_temporal_classification" }, { "name": "Deep Belief Networks", "org": "University of Toronto", "country": "Canada", "date": "2006-07-01", "precision": "month", "category": "historic", "open_weights": false, "params": "1.6M", "note": "Hinton's greedy layer-wise pretraining; kicked off the deep learning revival", "source": "https://en.wikipedia.org/wiki/Deep_belief_network" }, { "name": "Stupid Backoff n-gram LM ('Large Language Models in Machine Translation')", "org": "Google", "country": "USA", "date": "2007-06-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Trillion-token n-gram LM; early use of the phrase 'large language models'", "source": "https://www.semanticscholar.org/paper/Large-Language-Models-in-Machine-Translation-Brants-Popat/ba786c46373892554b98df42df7af6f5da343c9d" }, { "name": "Chinook (checkers solved)", "org": "University of Alberta", "country": "Canada", "date": "2007-07-19", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Science paper proves checkers is a draw with perfect play", "source": "https://en.wikipedia.org/wiki/Chinook_(computer_program)" }, { "name": "Polaris", "org": "University of Alberta", "country": "Canada", "date": "2007-07-23", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Limit hold'em poker bot; its 2008 version beat professional players", "source": "https://en.wikipedia.org/wiki/Polaris_(poker_bot)" }, { "name": "Boss", "org": "Carnegie Mellon University", "country": "USA", "date": "2007-11-03", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Tartan Racing's autonomous SUV; won DARPA Urban Challenge in traffic", "source": "https://en.wikipedia.org/wiki/DARPA_Grand_Challenge" }, { "name": "Collobert-Weston unified NLP network", "org": "NEC Labs America", "country": "USA", "date": "2008-07-01", "precision": "month", "category": "language", "open_weights": false, "params": "1.5M", "note": "Multitask deep net with learned word embeddings for many NLP tasks", "source": "https://dl.acm.org/doi/10.1145/1390156.1390177" }, { "name": "Google Voice Search (iPhone)", "org": "Google", "country": "USA", "date": "2008-11-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Brought cloud speech recognition to smartphones via Google Mobile App", "source": "https://en.wikipedia.org/wiki/Google_Voice_Search" }, { "name": "Deep Belief Network acoustic model", "org": "University of Toronto", "country": "Canada", "date": "2009-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "18M", "note": "Mohamed, Dahl & Hinton's deep nets for phone recognition; speech DL breakthrough", "source": "https://www.cs.utoronto.ca/~gdahl/papers/dbnPhoneRec.pdf" }, { "name": "IBM Watson (DeepQA)", "org": "IBM", "country": "USA", "date": "2009-04-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Announced Apr 2009; beat Jeopardy! champions Jennings and Rutter Feb 2011", "source": "https://en.wikipedia.org/wiki/IBM_Watson" }, { "name": "Adam (Robot Scientist)", "org": "Aberystwyth University", "country": "UK", "date": "2009-04-02", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "First machine to independently discover new scientific knowledge (yeast genetics)", "source": "https://en.wikipedia.org/wiki/Robot_Scientist" }, { "name": "GPU-trained Deep Belief Network (Raina, Madhavan & Ng)", "org": "Stanford University", "country": "USA", "date": "2009-06-01", "precision": "month", "category": "historic", "open_weights": false, "params": "100M", "note": "Showed GPUs make large-scale deep learning practical", "source": "https://dl.acm.org/doi/abs/10.1145/1553374.1553486" }, { "name": "Kinect (Project Natal)", "org": "Microsoft", "country": "USA", "date": "2009-06-01", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "ML body tracking from depth; announced E3 2009, shipped Nov 2010", "source": "https://en.wikipedia.org/wiki/Kinect" }, { "name": "CD-DNN-HMM speech recognizer", "org": "Microsoft", "country": "USA", "date": "2010-01-01", "precision": "year", "category": "audio", "open_weights": false, "params": "", "note": "Context-dependent deep nets gave large-vocabulary speech recognition breakthrough", "source": "https://en.wikipedia.org/wiki/Speech_recognition" }, { "name": "Siri (Siri Inc. app)", "org": "Siri Inc.", "country": "USA", "date": "2010-02-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "SRI spin-off voice assistant iPhone app; acquired by Apple April 2010", "source": "https://www.sri.com/hoi/siri/" }, { "name": "RNNLM (recurrent neural network language model)", "org": "Brno University of Technology", "country": "Czech Republic", "date": "2010-09-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Mikolov's RNN language model, substantially beating n-gram models", "source": "https://www.semanticscholar.org/paper/Recurrent-neural-network-based-language-model-Mikolov-Karafi%C3%A1t/9819b600a828a57e1cde047bbe710d3446b30da5" }, { "name": "Google self-driving car", "org": "Google", "country": "USA", "date": "2010-10-09", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Secret project since Jan 2009 revealed; precursor of Waymo", "source": "https://en.wikipedia.org/wiki/Waymo" }, { "name": "DanNet (GPU CNN)", "org": "IDSIA", "country": "Switzerland", "date": "2011-02-01", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Deep GPU CNNs winning vision contests; superhuman traffic-sign recognition 2011", "source": "https://arxiv.org/abs/1102.0183" }, { "name": "Siri (iPhone 4S)", "org": "Apple", "country": "USA", "date": "2011-10-04", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Voice assistant built into iPhone 4S; mainstreamed conversational AI", "source": "https://www.sri.com/hoi/siri/" }, { "name": "Google Brain \"cat\" network", "org": "Google", "country": "USA", "date": "2011-12-29", "precision": "day", "category": "vision", "open_weights": false, "params": "1B", "note": "16,000-core network learned cat and face detectors from unlabeled YouTube frames", "source": "https://arxiv.org/abs/1112.6209" }, { "name": "Multi-column Deep Neural Network (MCDNN)", "org": "IDSIA", "country": "Switzerland", "date": "2012-02-13", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "GPU CNN committees reached human-competitive accuracy on MNIST and traffic signs", "source": "https://arxiv.org/abs/1202.2745" }, { "name": "AlexNet", "org": "University of Toronto", "country": "Canada", "date": "2012-09-30", "precision": "day", "category": "vision", "open_weights": false, "params": "60M", "note": "GPU-trained CNN won ILSVRC 2012 by a wide margin, igniting deep learning", "source": "https://en.wikipedia.org/wiki/AlexNet" }, { "name": "word2vec", "org": "Google", "country": "USA", "date": "2013-01-16", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Efficient word embeddings famous for 'king - man + woman = queen' analogies", "source": "https://arxiv.org/abs/1301.3781" }, { "name": "Graves Handwriting Synthesis RNN", "org": "University of Toronto", "country": "Canada", "date": "2013-08-04", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "LSTM generating text and realistic handwriting with a soft attention window", "source": "https://arxiv.org/abs/1308.0850" }, { "name": "RCTM (Recurrent Continuous Translation Models)", "org": "University of Oxford", "country": "UK", "date": "2013-10-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Among the first end-to-end neural machine translation models", "source": "https://aclanthology.org/D13-1176/" }, { "name": "RNTN (Recursive Neural Tensor Network)", "org": "Stanford University", "country": "USA", "date": "2013-10-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Compositional sentiment model introduced with the Stanford Sentiment Treebank", "source": "https://aclanthology.org/D13-1170/" }, { "name": "R-CNN", "org": "UC Berkeley", "country": "USA", "date": "2013-11-11", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "CNN features on region proposals transformed object detection accuracy", "source": "https://arxiv.org/abs/1311.2524" }, { "name": "ZFNet", "org": "New York University", "country": "USA", "date": "2013-11-12", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Visualized CNN internals; basis of ILSVRC 2013 winning classification entry", "source": "https://arxiv.org/abs/1311.2901" }, { "name": "DeViSE", "org": "Google", "country": "USA", "date": "2013-12-01", "precision": "month", "category": "vision", "open_weights": false, "params": "", "note": "Joint image-word embedding enabling zero-shot recognition of unseen classes", "source": "https://www.semanticscholar.org/paper/DeViSE%3A-A-Deep-Visual-Semantic-Embedding-Model-Frome-Corrado/4aa4069693bee00d1b0759ca3df35e59284e9845" }, { "name": "Network in Network", "org": "National University of Singapore", "country": "Singapore", "date": "2013-12-16", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Introduced 1x1 convolution layers and global average pooling", "source": "https://arxiv.org/abs/1312.4400" }, { "name": "DQN (Deep Q-Network)", "org": "DeepMind", "country": "UK", "date": "2013-12-19", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "First deep RL agent learning Atari games directly from raw pixels", "source": "https://arxiv.org/abs/1312.5602" }, { "name": "Variational Autoencoder (VAE)", "org": "University of Amsterdam", "country": "Netherlands", "date": "2013-12-20", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Foundational latent-variable generative model using the reparameterization trick", "source": "https://arxiv.org/abs/1312.6114" }, { "name": "OverFeat", "org": "New York University", "country": "USA", "date": "2013-12-21", "precision": "day", "category": "vision", "open_weights": true, "params": "144M", "note": "Unified CNN for classification, localization, detection; won ILSVRC 2013 localization", "source": "https://arxiv.org/abs/1312.6229" }, { "name": "DeepFace", "org": "Meta", "country": "USA", "date": "2014-03-17", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Facebook face verification at ~97% on LFW, approaching human accuracy", "source": "https://www.technologyreview.com/2014/03/17/13822/facebook-creates-software-that-matches-faces-almost-as-well-as-you-do/" }, { "name": "Xiaoice", "org": "Microsoft", "country": "USA", "date": "2014-05-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Empathetic social chatbot launched in China via WeChat; went viral", "source": "https://zh.wikipedia.org/wiki/%E5%B0%8F%E5%86%B0" }, { "name": "Paragraph Vector (doc2vec)", "org": "Google", "country": "USA", "date": "2014-05-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Extended word2vec-style embeddings to sentences and documents", "source": "https://arxiv.org/abs/1405.4053" }, { "name": "RNN Encoder-Decoder (GRU)", "org": "Université de Montréal", "country": "Canada", "date": "2014-06-03", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Introduced the gated recurrent unit and RNN encoder-decoder for translation", "source": "https://arxiv.org/abs/1406.1078" }, { "name": "Generative Adversarial Networks (GAN)", "org": "Université de Montréal", "country": "Canada", "date": "2014-06-10", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Generator-versus-discriminator training launched adversarial generative modeling", "source": "https://arxiv.org/abs/1406.2661" }, { "name": "DeepID2", "org": "Chinese University of Hong Kong", "country": "China", "date": "2014-06-18", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Face features reaching 99.15% on LFW, surpassing human-level verification", "source": "https://arxiv.org/abs/1406.4773" }, { "name": "SPP-net", "org": "Microsoft", "country": "USA", "date": "2014-06-18", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "MSRA spatial pyramid pooling CNN; 24-102x faster detection than R-CNN", "source": "https://arxiv.org/abs/1406.4729" }, { "name": "GloVe", "org": "Stanford University", "country": "USA", "date": "2014-08-01", "precision": "month", "category": "language", "open_weights": true, "params": "", "note": "Global co-occurrence word vectors; widely used pretrained embeddings", "source": "https://nlp.stanford.edu/projects/glove/" }, { "name": "CNN for Sentence Classification (Kim CNN)", "org": "New York University", "country": "USA", "date": "2014-08-25", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Simple CNN over word vectors set text-classification state of the art", "source": "https://arxiv.org/abs/1408.5882" }, { "name": "RNNsearch (Bahdanau attention)", "org": "Université de Montréal", "country": "Canada", "date": "2014-09-01", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Introduced neural attention, letting translation models align source and target", "source": "https://arxiv.org/abs/1409.0473" }, { "name": "VGGNet (VGG-16/19)", "org": "University of Oxford", "country": "UK", "date": "2014-09-04", "precision": "day", "category": "vision", "open_weights": true, "params": "138M", "note": "Very deep 3x3-convolution networks; widely used pretrained backbone", "source": "https://arxiv.org/abs/1409.1556" }, { "name": "Seq2Seq LSTM", "org": "Google", "country": "USA", "date": "2014-09-10", "precision": "day", "category": "language", "open_weights": false, "params": "384M", "note": "Deep LSTM encoder-decoder established sequence-to-sequence learning for translation", "source": "https://arxiv.org/abs/1409.3215" }, { "name": "GoogLeNet (Inception v1)", "org": "Google", "country": "USA", "date": "2014-09-17", "precision": "day", "category": "vision", "open_weights": true, "params": "6.8M", "note": "22-layer Inception network won ILSVRC 2014 classification", "source": "https://arxiv.org/abs/1409.4842" }, { "name": "Memory Networks", "org": "Meta", "country": "USA", "date": "2014-10-15", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Networks with explicit long-term memory for question answering", "source": "https://arxiv.org/abs/1410.3916" }, { "name": "Neural Turing Machine", "org": "DeepMind", "country": "UK", "date": "2014-10-20", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Neural network with differentiable external memory learned simple algorithms", "source": "https://arxiv.org/abs/1410.5401" }, { "name": "Alexa", "org": "Amazon", "country": "USA", "date": "2014-11-06", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Voice assistant announced alongside the Echo smart speaker", "source": "https://en.wikipedia.org/wiki/Amazon_Alexa" }, { "name": "Fully Convolutional Networks (FCN)", "org": "UC Berkeley", "country": "USA", "date": "2014-11-14", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "End-to-end pixelwise semantic segmentation with fully convolutional networks", "source": "https://arxiv.org/abs/1411.4038" }, { "name": "Show and Tell", "org": "Google", "country": "USA", "date": "2014-11-17", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "CNN-plus-LSTM neural image caption generator", "source": "https://arxiv.org/abs/1411.4555" }, { "name": "Deep Speech", "org": "Baidu", "country": "China", "date": "2014-12-17", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "End-to-end RNN speech recognition outperforming traditional pipelines in noise", "source": "https://arxiv.org/abs/1412.5567" }, { "name": "DeepLab", "org": "Google", "country": "USA", "date": "2014-12-22", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Atrous convolution plus CRF for semantic image segmentation", "source": "https://arxiv.org/abs/1412.7062" }, { "name": "Cepheus", "org": "University of Alberta", "country": "Canada", "date": "2015-01-08", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Essentially solved heads-up limit Texas hold'em poker", "source": "https://doi.org/10.1126/science.1259433" }, { "name": "PReLU-Net", "org": "Microsoft", "country": "USA", "date": "2015-02-06", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "First reported to surpass human-level ImageNet error (4.94%)", "source": "https://arxiv.org/abs/1502.01852" }, { "name": "Show, Attend and Tell", "org": "Université de Montréal", "country": "Canada", "date": "2015-02-10", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Visual attention for image captioning; highly influential attention model", "source": "https://arxiv.org/abs/1502.03044" }, { "name": "DRAW", "org": "DeepMind", "country": "UK", "date": "2015-02-16", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Recurrent attention network generating images step by step", "source": "https://arxiv.org/abs/1502.04623" }, { "name": "Deep Q-Network (DQN)", "org": "DeepMind", "country": "UK", "date": "2015-02-25", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Human-level play across 49 Atari 2600 games learned from pixels", "source": "https://doi.org/10.1038/nature14236" }, { "name": "Diffusion Probabilistic Model", "org": "Stanford University", "country": "USA", "date": "2015-03-12", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Introduced diffusion models, later the basis of modern image generators", "source": "https://arxiv.org/abs/1503.03585" }, { "name": "FaceNet", "org": "Google", "country": "USA", "date": "2015-03-12", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Triplet-loss face embeddings reaching 99.63% on LFW", "source": "https://arxiv.org/abs/1503.03832" }, { "name": "End-to-End Visuomotor Policies", "org": "UC Berkeley", "country": "USA", "date": "2015-04-02", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Deep CNN policies mapping camera pixels directly to robot motor torques", "source": "https://arxiv.org/abs/1504.00702" }, { "name": "Claudico", "org": "Carnegie Mellon University", "country": "USA", "date": "2015-04-24", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "CMU no-limit hold'em AI; lost to top pros, precursor to Libratus", "source": "https://en.wikipedia.org/wiki/Claudico" }, { "name": "Fast R-CNN", "org": "Microsoft", "country": "USA", "date": "2015-04-30", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Faster, jointly trained region-based CNN object detector", "source": "https://arxiv.org/abs/1504.08083" }, { "name": "Highway Networks", "org": "IDSIA", "country": "Switzerland", "date": "2015-05-03", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Gated skip connections enabled training very deep nets; ResNet precursor", "source": "https://arxiv.org/abs/1505.00387" }, { "name": "U-Net", "org": "University of Freiburg", "country": "Germany", "date": "2015-05-18", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Encoder-decoder with skip connections; standard for biomedical segmentation", "source": "https://arxiv.org/abs/1505.04597" }, { "name": "char-rnn", "org": "Stanford University", "country": "USA", "date": "2015-05-21", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Karpathy's character-level LSTM text generator popularized neural text generation", "source": "https://karpathy.github.io/2015/05/21/rnn-effectiveness/" }, { "name": "Faster R-CNN", "org": "Microsoft", "country": "USA", "date": "2015-06-04", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Region Proposal Network enabled near real-time two-stage detection", "source": "https://arxiv.org/abs/1506.01497" }, { "name": "YOLO", "org": "University of Washington", "country": "USA", "date": "2015-06-08", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Single-pass, real-time object detection", "source": "https://arxiv.org/abs/1506.02640" }, { "name": "DeepDream", "org": "Google", "country": "USA", "date": "2015-06-17", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "'Inceptionism' visualizations turned CNN features into viral psychedelic imagery", "source": "https://research.google/blog/inceptionism-going-deeper-into-neural-networks/" }, { "name": "Neural Conversational Model", "org": "Google", "country": "USA", "date": "2015-06-19", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Seq2seq chatbot trained on movie subtitles and IT helpdesk logs", "source": "https://arxiv.org/abs/1506.05869" }, { "name": "Skip-Thought Vectors", "org": "University of Toronto", "country": "Canada", "date": "2015-06-22", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Unsupervised sentence embeddings learned by predicting neighboring sentences", "source": "https://arxiv.org/abs/1506.06726" }, { "name": "DeepBind", "org": "University of Toronto", "country": "Canada", "date": "2015-07-27", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Deep learning predicted DNA- and RNA-binding protein specificities", "source": "https://doi.org/10.1038/nbt.3300" }, { "name": "Listen, Attend and Spell", "org": "Google", "country": "USA", "date": "2015-08-05", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Attention-based end-to-end speech recognizer", "source": "https://arxiv.org/abs/1508.01211" }, { "name": "Neural Style Transfer", "org": "University of Tübingen", "country": "Germany", "date": "2015-08-26", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Rendered photos in famous painters' styles via CNN feature statistics", "source": "https://arxiv.org/abs/1508.06576" }, { "name": "Giraffe", "org": "Imperial College London", "country": "UK", "date": "2015-09-04", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Deep RL chess engine reaching roughly International Master strength", "source": "https://arxiv.org/abs/1509.01549" }, { "name": "DDPG (Deep Deterministic Policy Gradient)", "org": "DeepMind", "country": "UK", "date": "2015-09-09", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Deep RL for continuous control; foundational for robot learning", "source": "https://arxiv.org/abs/1509.02971" }, { "name": "AtomNet", "org": "Atomwise", "country": "USA", "date": "2015-10-10", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "First deep CNN for structure-based drug bioactivity prediction", "source": "https://arxiv.org/abs/1510.02855" }, { "name": "RankBrain", "org": "Google", "country": "USA", "date": "2015-10-26", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Machine-learning system interpreting Google Search queries", "source": "https://en.wikipedia.org/wiki/RankBrain" }, { "name": "Smart Reply", "org": "Google", "country": "USA", "date": "2015-11-03", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "LSTM model suggesting short automatic email replies in Inbox", "source": "https://research.google/blog/computer-respond-to-this-email/" }, { "name": "alignDRAW", "org": "University of Toronto", "country": "Canada", "date": "2015-11-09", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Early text-to-image model generating images from captions", "source": "https://arxiv.org/abs/1511.02793" }, { "name": "DCGAN", "org": "indico / Meta", "country": "USA", "date": "2015-11-19", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Convolutional GAN design that made GAN training stable and practical", "source": "https://arxiv.org/abs/1511.06434" }, { "name": "Inception v3", "org": "Google", "country": "USA", "date": "2015-12-02", "precision": "day", "category": "vision", "open_weights": true, "params": "23.6M", "note": "Factorized-convolution Inception; widely used pretrained image backbone", "source": "https://arxiv.org/abs/1512.00567" }, { "name": "Deep Speech 2", "org": "Baidu", "country": "China", "date": "2015-12-08", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "End-to-end English and Mandarin speech recognition at scale", "source": "https://arxiv.org/abs/1512.02595" }, { "name": "SSD (Single Shot MultiBox Detector)", "org": "UNC Chapel Hill", "country": "USA", "date": "2015-12-08", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Fast one-stage object detector", "source": "https://arxiv.org/abs/1512.02325" }, { "name": "ResNet", "org": "Microsoft", "country": "USA", "date": "2015-12-10", "precision": "day", "category": "vision", "open_weights": true, "params": "60.2M", "note": "Residual connections enabled 152-layer nets; won ILSVRC 2015", "source": "https://arxiv.org/abs/1512.03385" }, { "name": "PixelRNN", "org": "DeepMind", "country": "UK", "date": "2016-01-25", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Autoregressive pixel-by-pixel image generation; ICML 2016 best paper", "source": "https://arxiv.org/abs/1601.06759" }, { "name": "AlphaGo (AlphaGo Fan)", "org": "DeepMind", "country": "UK", "date": "2016-01-27", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "First program to beat a professional Go player without handicap", "source": "https://en.wikipedia.org/wiki/AlphaGo" }, { "name": "A3C (Asynchronous Advantage Actor-Critic)", "org": "DeepMind", "country": "UK", "date": "2016-02-04", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Asynchronous deep RL agent beating DQN on Atari using CPUs", "source": "https://arxiv.org/abs/1602.01783" }, { "name": "Big LSTM LM (lm_1b)", "org": "Google", "country": "USA", "date": "2016-02-07", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Scaled LSTM language models on the One Billion Word benchmark", "source": "https://arxiv.org/abs/1602.02410" }, { "name": "Inception-v4 / Inception-ResNet-v2", "org": "Google", "country": "USA", "date": "2016-02-23", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Combined Inception modules with residual connections", "source": "https://arxiv.org/abs/1602.07261" }, { "name": "SqueezeNet", "org": "DeepScale / UC Berkeley", "country": "USA", "date": "2016-02-24", "precision": "day", "category": "vision", "open_weights": true, "params": "1.2M", "note": "AlexNet-level accuracy with 50x fewer parameters", "source": "https://arxiv.org/abs/1602.07360" }, { "name": "Face2Face", "org": "Technical University of Munich", "country": "Germany", "date": "2016-03-01", "precision": "month", "category": "video", "open_weights": false, "params": "", "note": "Real-time facial reenactment of video; an early deepfake precursor", "source": "https://niessnerlab.org/projects/thies2016face.html" }, { "name": "Google Robotic Grasping (Hand-Eye Coordination)", "org": "Google", "country": "USA", "date": "2016-03-07", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Learned grasping from over 800,000 attempts across many robot arms", "source": "https://arxiv.org/abs/1603.02199" }, { "name": "AlphaGo Lee", "org": "DeepMind", "country": "UK", "date": "2016-03-09", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Beat Lee Sedol 4-1 in Seoul; watershed moment for AI", "source": "https://en.wikipedia.org/wiki/AlphaGo" }, { "name": "Tay", "org": "Microsoft", "country": "USA", "date": "2016-03-23", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Twitter chatbot pulled within 16 hours after posting offensive tweets", "source": "https://en.wikipedia.org/wiki/Tay_(chatbot)" }, { "name": "Fast Neural Style Transfer (Perceptual Losses)", "org": "Stanford University", "country": "USA", "date": "2016-03-27", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Feed-forward real-time style transfer; enabled mobile style apps", "source": "https://arxiv.org/abs/1603.08155" }, { "name": "PilotNet (DAVE-2)", "org": "NVIDIA", "country": "USA", "date": "2016-04-25", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "End-to-end CNN steering a car directly from camera pixels", "source": "https://arxiv.org/abs/1604.07316" }, { "name": "SyntaxNet (Parsey McParseface)", "org": "Google", "country": "USA", "date": "2016-05-12", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Open-sourced neural parser billed as world's most accurate", "source": "https://research.google/blog/announcing-syntaxnet-the-worlds-most-accurate-parser-goes-open-source/" }, { "name": "GAN-INT-CLS (Text-to-Image GAN)", "org": "University of Michigan", "country": "USA", "date": "2016-05-17", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "GAN synthesizing plausible images from natural-language descriptions", "source": "https://arxiv.org/abs/1605.05396" }, { "name": "Prisma", "org": "Prisma Labs", "country": "Russia", "date": "2016-06-11", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Viral app bringing neural style transfer to smartphones", "source": "https://en.wikipedia.org/wiki/Prisma_(app)" }, { "name": "PixelCNN", "org": "DeepMind", "country": "UK", "date": "2016-06-16", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Gated convolutional autoregressive image generator; basis for WaveNet", "source": "https://arxiv.org/abs/1606.05328" }, { "name": "fastText", "org": "Meta", "country": "USA", "date": "2016-07-06", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Fast text classification and subword word vectors, open-sourced", "source": "https://arxiv.org/abs/1607.01759" }, { "name": "DenseNet", "org": "Cornell University", "country": "USA", "date": "2016-08-25", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Densely connected convolutional layers; CVPR 2017 best paper", "source": "https://arxiv.org/abs/1608.06993" }, { "name": "RaptorX-Contact", "org": "Toyota Technological Institute at Chicago", "country": "USA", "date": "2016-09-02", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Ultra-deep ResNet protein contact prediction; forerunner of AlphaFold", "source": "https://arxiv.org/abs/1609.00680" }, { "name": "VGAN (Generating Videos with Scene Dynamics)", "org": "MIT", "country": "USA", "date": "2016-09-08", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Early GAN generating short video clips with scene dynamics", "source": "https://arxiv.org/abs/1609.02612" }, { "name": "WaveNet", "org": "DeepMind", "country": "UK", "date": "2016-09-08", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Autoregressive raw-audio model producing far more natural synthetic speech", "source": "https://deepmind.google/discover/blog/wavenet-a-generative-model-for-raw-audio/" }, { "name": "SRGAN", "org": "Twitter", "country": "USA", "date": "2016-09-15", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "GAN-based photo-realistic 4x single-image super-resolution", "source": "https://arxiv.org/abs/1609.04802" }, { "name": "Google Neural Machine Translation (GNMT)", "org": "Google", "country": "USA", "date": "2016-09-26", "precision": "day", "category": "language", "open_weights": false, "params": "278M", "note": "Neural MT replaced phrase-based system in Google Translate", "source": "https://arxiv.org/abs/1609.08144" }, { "name": "Chemical VAE (Automatic Chemical Design)", "org": "Harvard University", "country": "USA", "date": "2016-10-07", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Variational autoencoder generating and optimizing molecules in latent space", "source": "https://arxiv.org/abs/1610.02415" }, { "name": "Xception", "org": "Google", "country": "USA", "date": "2016-10-07", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Depthwise separable convolutions taking Inception to its extreme", "source": "https://arxiv.org/abs/1610.02357" }, { "name": "Differentiable Neural Computer", "org": "DeepMind", "country": "UK", "date": "2016-10-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Memory-augmented network solving graph tasks like London Underground routing", "source": "https://doi.org/10.1038/nature20101" }, { "name": "Microsoft Conversational Speech Recognition (Human Parity)", "org": "Microsoft", "country": "USA", "date": "2016-10-17", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "First claimed human parity on Switchboard conversational speech (5.9% WER)", "source": "https://arxiv.org/abs/1610.05256" }, { "name": "ByteNet", "org": "DeepMind", "country": "UK", "date": "2016-10-31", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Dilated convolutional translation model running in linear time", "source": "https://arxiv.org/abs/1610.10099" }, { "name": "Adobe VoCo", "org": "Adobe", "country": "USA", "date": "2016-11-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "'Photoshop for voice' speech-editing demo at Adobe MAX; never shipped", "source": "https://en.wikipedia.org/wiki/Adobe_Voco" }, { "name": "BiDAF", "org": "Allen Institute for AI", "country": "USA", "date": "2016-11-05", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Bidirectional attention flow model; state of the art on SQuAD", "source": "https://arxiv.org/abs/1611.01603" }, { "name": "LipNet", "org": "University of Oxford", "country": "UK", "date": "2016-11-05", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "End-to-end sentence-level lipreading from video", "source": "https://arxiv.org/abs/1611.01599" }, { "name": "Neural Architecture Search (NAS)", "org": "Google", "country": "USA", "date": "2016-11-05", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "RL controller automatically designed neural network architectures", "source": "https://arxiv.org/abs/1611.01578" }, { "name": "DeepCoder", "org": "Microsoft", "country": "USA", "date": "2016-11-07", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Neural-guided program synthesis writing short programs from examples", "source": "https://arxiv.org/abs/1611.01989" }, { "name": "Google Multilingual NMT (Zero-Shot Translation)", "org": "Google", "country": "USA", "date": "2016-11-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Single multilingual model translating between unseen language pairs", "source": "https://arxiv.org/abs/1611.04558" }, { "name": "ResNeXt", "org": "Meta", "country": "USA", "date": "2016-11-16", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Aggregated residual transformations; runner-up in ILSVRC 2016 classification", "source": "https://arxiv.org/abs/1611.05431" }, { "name": "DeepZenGo", "org": "Dwango", "country": "Japan", "date": "2016-11-19", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Japanese deep-learning Go AI; won one game against Cho Chikun", "source": "https://en.wikipedia.org/wiki/Zen_(software)" }, { "name": "pix2pix", "org": "UC Berkeley", "country": "USA", "date": "2016-11-21", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Conditional GAN framework for general image-to-image translation", "source": "https://arxiv.org/abs/1611.07004" }, { "name": "OpenPose (Part Affinity Fields)", "org": "Carnegie Mellon University", "country": "USA", "date": "2016-11-24", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Real-time multi-person 2D pose estimation; became OpenPose library", "source": "https://arxiv.org/abs/1611.08050" }, { "name": "Google Diabetic Retinopathy Model", "org": "Google", "country": "USA", "date": "2016-11-29", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Detected diabetic eye disease from retinal photos at ophthalmologist level", "source": "https://research.google/blog/deep-learning-for-detection-of-diabetic-eye-disease/" }, { "name": "Zo", "org": "Microsoft", "country": "USA", "date": "2016-12-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Social chatbot successor to Tay, launched on Kik Messenger", "source": "https://en.wikipedia.org/wiki/Zo_(bot)" }, { "name": "PointNet", "org": "Stanford University", "country": "USA", "date": "2016-12-02", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Deep learning directly on raw 3D point clouds", "source": "https://arxiv.org/abs/1612.00593" }, { "name": "StackGAN", "org": "Rutgers University", "country": "USA", "date": "2016-12-10", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Two-stage GAN producing 256x256 photo-realistic images from text", "source": "https://arxiv.org/abs/1612.03242" }, { "name": "DeepVariant", "org": "Google", "country": "USA", "date": "2016-12-14", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "CNN genetic variant caller treating sequencing reads as images", "source": "https://doi.org/10.1101/092890" }, { "name": "SampleRNN", "org": "Université de Montréal", "country": "Canada", "date": "2016-12-22", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Hierarchical RNN generating raw audio unconditionally, incl. music", "source": "https://arxiv.org/abs/1612.07837" }, { "name": "YOLOv2 (YOLO9000)", "org": "University of Washington", "country": "USA", "date": "2016-12-25", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Faster, more accurate YOLO detecting over 9,000 object categories", "source": "https://arxiv.org/abs/1612.08242" }, { "name": "AlphaGo Master", "org": "DeepMind", "country": "UK", "date": "2016-12-29", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Went 60-0 online against top pros; later beat Ke Jie 3-0", "source": "https://en.wikipedia.org/wiki/AlphaGo" }, { "name": "Libratus", "org": "Carnegie Mellon University", "country": "USA", "date": "2017-01-04", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Beat four top professionals at heads-up no-limit hold'em", "source": "https://en.wikipedia.org/wiki/Libratus" }, { "name": "DeepStack", "org": "University of Alberta", "country": "Canada", "date": "2017-01-06", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Beat professional players at heads-up no-limit hold'em using deep learning", "source": "https://arxiv.org/abs/1701.01724" }, { "name": "Sparsely-Gated Mixture-of-Experts (MoE)", "org": "Google", "country": "USA", "date": "2017-01-23", "precision": "day", "category": "language", "open_weights": false, "params": "137B (largest variant)", "note": "Up to 137B-parameter sparse MoE; forerunner of modern MoE LLMs", "source": "https://arxiv.org/abs/1701.06538" }, { "name": "Dermatologist-Level Skin Cancer CNN", "org": "Stanford University", "country": "USA", "date": "2017-01-25", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Dermatologist-level skin cancer classification from photographs", "source": "https://doi.org/10.1038/nature21056" }, { "name": "Wasserstein GAN (WGAN)", "org": "Meta", "country": "USA", "date": "2017-01-26", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Wasserstein loss that stabilized GAN training", "source": "https://arxiv.org/abs/1701.07875" }, { "name": "Deep Voice", "org": "Baidu", "country": "China", "date": "2017-02-25", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Real-time text-to-speech built entirely from deep neural networks", "source": "https://arxiv.org/abs/1702.07825" }, { "name": "Fine Art (Jueyi)", "org": "Tencent", "country": "China", "date": "2017-03-18", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Tencent's Go AI; won the 10th UEC Cup computer Go tournament", "source": "https://technode.com/2017/03/20/tencents-fine-art-wins-computer-go-uec-cup/" }, { "name": "Mask R-CNN", "org": "Meta", "country": "USA", "date": "2017-03-20", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Instance segmentation extending Faster R-CNN; ICCV 2017 best paper", "source": "https://arxiv.org/abs/1703.06870" }, { "name": "Tacotron", "org": "Google", "country": "USA", "date": "2017-03-29", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "End-to-end text-to-speech from characters to spectrograms", "source": "https://arxiv.org/abs/1703.10135" }, { "name": "CycleGAN", "org": "UC Berkeley", "country": "USA", "date": "2017-03-30", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Unpaired image-to-image translation, e.g., horses into zebras", "source": "https://arxiv.org/abs/1703.10593" }, { "name": "Lyrebird", "org": "Lyrebird", "country": "Canada", "date": "2017-04-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Montreal startup's voice cloning from about one minute of audio", "source": "https://www.scientificamerican.com/article/new-ai-tech-can-mimic-any-voice/" }, { "name": "MPNN (Neural Message Passing for Quantum Chemistry)", "org": "Google", "country": "USA", "date": "2017-04-04", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Graph neural networks predicting molecular quantum-chemical properties", "source": "https://arxiv.org/abs/1704.01212" }, { "name": "NSynth", "org": "Google", "country": "USA", "date": "2017-04-05", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "WaveNet autoencoder synthesizing novel musical instrument sounds", "source": "https://arxiv.org/abs/1704.01279" }, { "name": "Sentiment Neuron", "org": "OpenAI", "country": "USA", "date": "2017-04-05", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Byte-level LSTM on Amazon reviews developed an unsupervised sentiment neuron", "source": "https://arxiv.org/abs/1704.01444" }, { "name": "Sketch-RNN", "org": "Google", "country": "USA", "date": "2017-04-11", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Generative recurrent model of vector sketches from Quick, Draw! data", "source": "https://arxiv.org/abs/1704.03477" }, { "name": "MobileNet", "org": "Google", "country": "USA", "date": "2017-04-17", "precision": "day", "category": "vision", "open_weights": true, "params": "4.2M", "note": "Efficient depthwise-separable CNNs designed for mobile devices", "source": "https://arxiv.org/abs/1704.04861" }, { "name": "InferSent", "org": "Meta", "country": "USA", "date": "2017-05-05", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Universal sentence embeddings trained on natural language inference data", "source": "https://arxiv.org/abs/1705.02364" }, { "name": "ConvS2S", "org": "Meta", "country": "USA", "date": "2017-05-08", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Fully convolutional seq2seq translation, faster than recurrent models", "source": "https://arxiv.org/abs/1705.03122" }, { "name": "pix2code", "org": "Uizard", "country": "Denmark", "date": "2017-05-22", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Generated GUI code directly from interface screenshots", "source": "https://arxiv.org/abs/1705.07962" }, { "name": "Deep Voice 2", "org": "Baidu", "country": "China", "date": "2017-05-24", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Multi-speaker neural text-to-speech learning hundreds of voices", "source": "https://arxiv.org/abs/1705.08947" }, { "name": "Transformer", "org": "Google", "country": "USA", "date": "2017-06-12", "precision": "day", "category": "language", "open_weights": false, "params": "213M", "note": "'Attention Is All You Need'; architecture behind nearly all modern LLMs", "source": "https://arxiv.org/abs/1706.03762" }, { "name": "Hybrid Reward Architecture (Ms. Pac-Man)", "org": "Microsoft", "country": "USA", "date": "2017-06-13", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Maluuba agent reached the maximum 999,990 score in Ms. Pac-Man", "source": "https://arxiv.org/abs/1706.04208" }, { "name": "Deal or No Deal Negotiation Agents", "org": "Meta", "country": "USA", "date": "2017-06-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "End-to-end negotiation dialogue agents; misreported as inventing language", "source": "https://engineering.fb.com/2017/06/14/ml-applications/deal-or-no-deal-training-ai-bots-to-negotiate/" }, { "name": "MultiModel (One Model To Learn Them All)", "org": "Google", "country": "USA", "date": "2017-06-16", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Single model trained jointly on image, speech, translation and parsing", "source": "https://arxiv.org/abs/1706.05137" }, { "name": "DeepLabv3", "org": "Google", "country": "USA", "date": "2017-06-17", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Atrous spatial pyramid pooling for state-of-the-art semantic segmentation", "source": "https://arxiv.org/abs/1706.05587" }, { "name": "SchNet", "org": "TU Berlin", "country": "Germany", "date": "2017-06-26", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Continuous-filter convolutions modeling quantum interactions in molecules", "source": "https://arxiv.org/abs/1706.08566" }, { "name": "ShuffleNet", "org": "Megvii", "country": "China", "date": "2017-07-04", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Channel-shuffle CNN for extremely efficient mobile inference", "source": "https://arxiv.org/abs/1707.01083" }, { "name": "Synthesizing Obama", "org": "University of Washington", "country": "USA", "date": "2017-07-11", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Generated realistic lip-synced Obama video from audio clips", "source": "https://www.washington.edu/news/2017/07/11/lip-syncing-obama-new-tools-turn-audio-clips-into-realistic-video/" }, { "name": "NASNet", "org": "Google", "country": "USA", "date": "2017-07-21", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Architecture-search-designed CNN achieving state-of-the-art ImageNet accuracy", "source": "https://arxiv.org/abs/1707.07012" }, { "name": "CoVe (Contextualized Word Vectors)", "org": "Salesforce", "country": "USA", "date": "2017-08-01", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Contextualized word vectors from a translation encoder; precursor to ELMo", "source": "https://arxiv.org/abs/1708.00107" }, { "name": "RetinaNet", "org": "Meta", "country": "USA", "date": "2017-08-07", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Focal-loss one-stage detector matching two-stage detector accuracy", "source": "https://arxiv.org/abs/1708.02002" }, { "name": "OpenAI Dota 2 1v1 Bot", "org": "OpenAI", "country": "USA", "date": "2017-08-11", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Self-play bot beat pro Dendi at 1v1 during The International 2017", "source": "https://en.wikipedia.org/wiki/OpenAI_Five" }, { "name": "DeepL Translator", "org": "DeepL", "country": "Germany", "date": "2017-08-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Neural translator launched claiming quality beyond Google Translate", "source": "https://en.wikipedia.org/wiki/DeepL_Translator" }, { "name": "SENet (Squeeze-and-Excitation Networks)", "org": "Momenta", "country": "China", "date": "2017-09-05", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Squeeze-and-excitation blocks; won the final ILSVRC 2017 classification", "source": "https://arxiv.org/abs/1709.01507" }, { "name": "Parallel WaveNet", "org": "DeepMind", "country": "UK", "date": "2017-10-04", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "1,000x faster WaveNet launched for Google Assistant voices", "source": "https://deepmind.google/discover/blog/wavenet-launches-in-the-google-assistant/" }, { "name": "Rainbow", "org": "DeepMind", "country": "UK", "date": "2017-10-06", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Combined six DQN improvements into state-of-the-art Atari agent", "source": "https://arxiv.org/abs/1710.02298" }, { "name": "AlphaGo Zero", "org": "DeepMind", "country": "UK", "date": "2017-10-18", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Learned Go purely from self-play; beat AlphaGo Lee 100-0", "source": "https://deepmind.google/discover/blog/alphago-zero-starting-from-scratch/" }, { "name": "Deep Voice 3", "org": "Baidu", "country": "China", "date": "2017-10-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Fully convolutional TTS scaled to over 2,000 speakers", "source": "https://arxiv.org/abs/1710.07654" }, { "name": "Leela Zero", "org": "Gian-Carlo Pascutto", "country": "Belgium", "date": "2017-10-25", "precision": "day", "category": "games", "open_weights": true, "params": "", "note": "Open-source community replication of AlphaGo Zero", "source": "https://en.wikipedia.org/wiki/Leela_Zero" }, { "name": "Capsule Networks (CapsNet)", "org": "Google", "country": "USA", "date": "2017-10-26", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Hinton's capsule networks with dynamic routing between capsules", "source": "https://arxiv.org/abs/1710.09829" }, { "name": "Progressive GAN (ProGAN)", "org": "NVIDIA", "country": "USA", "date": "2017-10-27", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Progressive growing produced unprecedented 1024x1024 GAN-generated faces", "source": "https://arxiv.org/abs/1710.10196" }, { "name": "VQ-VAE", "org": "DeepMind", "country": "UK", "date": "2017-11-02", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Discrete-latent autoencoder later central to image and audio tokenizers", "source": "https://arxiv.org/abs/1711.00937" }, { "name": "CheXNet", "org": "Stanford University", "country": "USA", "date": "2017-11-14", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "121-layer DenseNet detecting pneumonia from chest X-rays at radiologist level", "source": "https://arxiv.org/abs/1711.05225" }, { "name": "StarGAN", "org": "NAVER Clova / Korea University", "country": "South Korea", "date": "2017-11-24", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Single GAN handling multi-domain facial attribute translation", "source": "https://arxiv.org/abs/1711.09020" }, { "name": "pix2pixHD", "org": "NVIDIA", "country": "USA", "date": "2017-11-30", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Conditional GAN synthesizing 2048x1024 images from semantic label maps", "source": "https://arxiv.org/abs/1711.11585" }, { "name": "AlphaZero", "org": "DeepMind", "country": "UK", "date": "2017-12-05", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Mastered chess, shogi and Go through self-play within hours", "source": "https://arxiv.org/abs/1712.01815" }, { "name": "Kepler Exoplanet CNN (Shallue & Vanderburg)", "org": "Google", "country": "USA", "date": "2017-12-14", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Neural network found Kepler-90i, revealing an eight-planet system", "source": "https://en.wikipedia.org/wiki/Kepler-90i" }, { "name": "Tacotron 2", "org": "Google", "country": "USA", "date": "2017-12-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Tacotron with WaveNet vocoder approaching human speech naturalness", "source": "https://arxiv.org/abs/1712.05884" }, { "name": "ULMFiT", "org": "fast.ai", "country": "USA", "date": "2018-01-18", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Popularized fine-tuning pretrained language models for NLP transfer learning", "source": "https://arxiv.org/abs/1801.06146" }, { "name": "ArcFace", "org": "Imperial College London", "country": "UK", "date": "2018-01-23", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Additive angular margin loss; became the standard face recognition model", "source": "https://arxiv.org/abs/1801.07698" }, { "name": "AmoebaNet", "org": "Google", "country": "USA", "date": "2018-02-05", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Evolution-discovered image classifier architecture setting ImageNet state of the art", "source": "https://arxiv.org/abs/1802.01548" }, { "name": "IMPALA", "org": "DeepMind", "country": "UK", "date": "2018-02-05", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Scalable distributed actor-learner RL agent mastering 30 DMLab tasks", "source": "https://arxiv.org/abs/1802.01561" }, { "name": "DeepLabv3+", "org": "Google", "country": "USA", "date": "2018-02-07", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Encoder-decoder atrous convolution model for semantic image segmentation", "source": "https://arxiv.org/abs/1802.02611" }, { "name": "ELMo", "org": "Allen Institute for AI", "country": "USA", "date": "2018-02-15", "precision": "day", "category": "language", "open_weights": true, "params": "94M", "note": "Deep contextual word embeddings; kicked off the NLP pretraining era", "source": "https://arxiv.org/abs/1802.05365" }, { "name": "WaveRNN", "org": "DeepMind", "country": "UK", "date": "2018-02-23", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Compact recurrent vocoder enabling real-time neural speech synthesis on devices", "source": "https://arxiv.org/abs/1802.08435" }, { "name": "MusicVAE", "org": "Google", "country": "USA", "date": "2018-03-13", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Magenta hierarchical VAE for interpolating and generating musical sequences", "source": "https://arxiv.org/abs/1803.05428" }, { "name": "World Models", "org": "Google", "country": "USA", "date": "2018-03-27", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Ha and Schmidhuber: agent learns inside its own generative dream", "source": "https://arxiv.org/abs/1803.10122" }, { "name": "Universal Sentence Encoder", "org": "Google", "country": "USA", "date": "2018-03-29", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "General-purpose sentence embeddings released on TensorFlow Hub", "source": "https://arxiv.org/abs/1803.11175" }, { "name": "YOLOv3", "org": "University of Washington", "country": "USA", "date": "2018-04-08", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Fast single-stage detector; ubiquitous real-time object detection baseline", "source": "https://arxiv.org/abs/1804.02767" }, { "name": "ELF OpenGo", "org": "Meta", "country": "USA", "date": "2018-05-02", "precision": "day", "category": "games", "open_weights": true, "params": "", "note": "Open-source AlphaZero-style Go bot that beat top professional players", "source": "https://research.facebook.com/blog/2018/05/facebook-open-sources-elf-opengo/" }, { "name": "ResNeXt WSL", "org": "Meta", "country": "USA", "date": "2018-05-02", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Facebook AI: ResNeXt weakly supervised on 3.5B Instagram images", "source": "https://arxiv.org/abs/1805.00932" }, { "name": "Google Duplex", "org": "Google", "country": "USA", "date": "2018-05-08", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Humanlike voice AI that phones businesses to book appointments", "source": "https://research.google/blog/google-duplex-an-ai-system-for-accomplishing-real-world-tasks-over-the-phone/" }, { "name": "SAGAN", "org": "Google", "country": "USA", "date": "2018-05-21", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Self-attention GAN; big jump in ImageNet image generation quality", "source": "https://arxiv.org/abs/1805.08318" }, { "name": "GPT-1", "org": "OpenAI", "country": "USA", "date": "2018-06-11", "precision": "day", "category": "language", "open_weights": true, "params": "117M", "note": "First generative pre-trained transformer: unsupervised pretraining plus fine-tuning", "source": "https://openai.com/blog/language-unsupervised/" }, { "name": "SV2TTS", "org": "Google", "country": "USA", "date": "2018-06-12", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Multispeaker TTS cloning unseen voices from seconds of audio", "source": "https://arxiv.org/abs/1806.04558" }, { "name": "Generative Query Network (GQN)", "org": "DeepMind", "country": "UK", "date": "2018-06-14", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Learns 3D scene representations and renders unseen viewpoints", "source": "https://deepmind.google/discover/blog/neural-scene-representation-and-rendering/" }, { "name": "Project Debater", "org": "IBM", "country": "USA", "date": "2018-06-18", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "First AI system to debate human champions live on complex topics", "source": "https://en.wikipedia.org/wiki/Project_Debater" }, { "name": "OpenAI Five", "org": "OpenAI", "country": "USA", "date": "2018-06-25", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Self-play RL team of five bots beating human Dota 2 teams", "source": "https://openai.com/blog/openai-five/" }, { "name": "QT-Opt", "org": "Google", "country": "USA", "date": "2018-06-27", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Large-scale RL for vision-based grasping on real robot arms", "source": "https://arxiv.org/abs/1806.10293" }, { "name": "FTW (Capture the Flag agent)", "org": "DeepMind", "country": "UK", "date": "2018-07-03", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Human-level team play in Quake III Arena Capture the Flag", "source": "https://arxiv.org/abs/1807.01281" }, { "name": "Glow", "org": "OpenAI", "country": "USA", "date": "2018-07-09", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Reversible flow-based generative model for high-resolution faces", "source": "https://openai.com/blog/glow/" }, { "name": "Universal Transformer", "org": "Google", "country": "USA", "date": "2018-07-10", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Recurrent-in-depth Transformer with adaptive computation time", "source": "https://arxiv.org/abs/1807.03819" }, { "name": "Dactyl", "org": "OpenAI", "country": "USA", "date": "2018-07-30", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Sim-trained robot hand manipulates objects with humanlike dexterity", "source": "https://openai.com/blog/learning-dexterity/" }, { "name": "vid2vid", "org": "NVIDIA", "country": "USA", "date": "2018-08-20", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "High-resolution video-to-video synthesis with conditional GANs", "source": "https://arxiv.org/abs/1808.06601" }, { "name": "Everybody Dance Now", "org": "UC Berkeley", "country": "USA", "date": "2018-08-22", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Motion transfer making anyone appear to dance like a professional", "source": "https://arxiv.org/abs/1808.07371" }, { "name": "DLSS", "org": "NVIDIA", "country": "USA", "date": "2018-09-01", "precision": "month", "category": "image", "open_weights": false, "params": "", "note": "Neural upscaling for real-time game rendering, launched with GeForce RTX", "source": "https://en.wikipedia.org/wiki/Deep_Learning_Super_Sampling" }, { "name": "ESRGAN", "org": "The Chinese University of Hong Kong", "country": "China", "date": "2018-09-01", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Enhanced GAN super-resolution; widely used for photo and game upscaling", "source": "https://arxiv.org/abs/1809.00219" }, { "name": "Music Transformer", "org": "Google", "country": "USA", "date": "2018-09-12", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Relative-attention transformer composing long coherent piano music", "source": "https://arxiv.org/abs/1809.04281" }, { "name": "BigGAN", "org": "DeepMind", "country": "UK", "date": "2018-09-28", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Large-scale class-conditional GAN; leap in ImageNet image fidelity", "source": "https://arxiv.org/abs/1809.11096" }, { "name": "BERT", "org": "Google", "country": "USA", "date": "2018-10-11", "precision": "day", "category": "language", "open_weights": true, "params": "340M", "note": "Bidirectional transformer pretraining; new state of the art across NLP", "source": "https://arxiv.org/abs/1810.04805" }, { "name": "Random Network Distillation", "org": "OpenAI", "country": "USA", "date": "2018-10-30", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Curiosity-driven agent first to exceed human average on Montezuma's Revenge", "source": "https://arxiv.org/abs/1810.12894" }, { "name": "WaveGlow", "org": "NVIDIA", "country": "USA", "date": "2018-10-31", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Flow-based vocoder for fast, high-quality speech synthesis", "source": "https://arxiv.org/abs/1811.00002" }, { "name": "GPipe", "org": "Google", "country": "USA", "date": "2018-11-16", "precision": "day", "category": "vision", "open_weights": false, "params": "557M", "note": "Pipeline parallelism trained a record 557M-parameter AmoebaNet", "source": "https://arxiv.org/abs/1811.06965" }, { "name": "Go-Explore", "org": "Uber", "country": "USA", "date": "2018-11-26", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Uber AI Labs: crushed notoriously hard-exploration Montezuma's Revenge", "source": "https://eng.uber.com/go-explore/" }, { "name": "AlphaFold", "org": "DeepMind", "country": "UK", "date": "2018-12-02", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Won CASP13 protein structure prediction; first AlphaFold", "source": "https://en.wikipedia.org/wiki/AlphaFold" }, { "name": "SlowFast", "org": "Meta", "country": "USA", "date": "2018-12-10", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Two-pathway network setting state of the art in video recognition", "source": "https://arxiv.org/abs/1812.03982" }, { "name": "StyleGAN", "org": "NVIDIA", "country": "USA", "date": "2018-12-12", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Style-based GAN with photorealistic faces; powered ThisPersonDoesNotExist", "source": "https://arxiv.org/abs/1812.04948" }, { "name": "LASER", "org": "Meta", "country": "USA", "date": "2018-12-26", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Massively multilingual sentence embeddings covering 93 languages", "source": "https://arxiv.org/abs/1812.10464" }, { "name": "Transformer-XL", "org": "Google", "country": "USA", "date": "2019-01-09", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Segment recurrence for long-context language modeling (with CMU)", "source": "https://arxiv.org/abs/1901.02860" }, { "name": "XLM", "org": "Meta", "country": "USA", "date": "2019-01-22", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Facebook AI cross-lingual language model pretraining", "source": "https://arxiv.org/abs/1901.07291" }, { "name": "AlphaStar", "org": "DeepMind", "country": "UK", "date": "2019-01-24", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Beat top StarCraft II professionals TLO and MaNa", "source": "https://www.deepmind.com/blog/alphastar-mastering-the-real-time-strategy-game-starcraft-ii" }, { "name": "BioBERT", "org": "Korea University", "country": "South Korea", "date": "2019-01-25", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "BERT pretrained on biomedical literature; landmark domain-specific language model", "source": "https://arxiv.org/abs/1901.08746" }, { "name": "Evolved Transformer", "org": "Google", "country": "USA", "date": "2019-01-30", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Transformer variant discovered by neural architecture search", "source": "https://arxiv.org/abs/1901.11117" }, { "name": "MT-DNN", "org": "Microsoft", "country": "USA", "date": "2019-01-31", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Multi-task BERT fine-tuning that topped the GLUE leaderboard", "source": "https://arxiv.org/abs/1901.11504" }, { "name": "GPT-2", "org": "OpenAI", "country": "USA", "date": "2019-02-14", "precision": "day", "category": "language", "open_weights": true, "params": "1.5B (124M released)", "note": "Full model withheld over misuse fears; staged release began", "source": "https://en.wikipedia.org/wiki/GPT-2" }, { "name": "KataGo", "org": "Jane Street", "country": "USA", "date": "2019-02-27", "precision": "day", "category": "games", "open_weights": true, "params": "", "note": "Open-source Go engine accelerating AlphaZero-style self-play roughly 50x", "source": "https://arxiv.org/abs/1902.10565" }, { "name": "ERNIE", "org": "Baidu", "country": "China", "date": "2019-03-01", "precision": "month", "category": "language", "open_weights": true, "params": "", "note": "Knowledge-enhanced BERT-style model; start of Baidu's ERNIE line", "source": "https://syncedreview.com/2019/03/25/baidus-ernie-tops-googles-bert-in-chinese-nlp-tasks/" }, { "name": "GauGAN", "org": "NVIDIA", "country": "USA", "date": "2019-03-18", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "SPADE model turning semantic sketches into photorealistic landscapes", "source": "https://arxiv.org/abs/1903.07291" }, { "name": "SciBERT", "org": "Allen Institute for AI", "country": "USA", "date": "2019-03-26", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "BERT pretrained on scientific papers with a science vocabulary", "source": "https://arxiv.org/abs/1903.10676" }, { "name": "VideoBERT", "org": "Google", "country": "USA", "date": "2019-04-03", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Joint BERT-style model over video tokens and speech transcripts", "source": "https://arxiv.org/abs/1904.01766" }, { "name": "Jasper", "org": "NVIDIA", "country": "USA", "date": "2019-04-05", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Deep all-convolutional end-to-end speech recognition model", "source": "https://arxiv.org/abs/1904.03288" }, { "name": "wav2vec", "org": "Meta", "country": "USA", "date": "2019-04-11", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Self-supervised pretraining on raw audio for speech recognition", "source": "https://arxiv.org/abs/1904.05862" }, { "name": "Translatotron", "org": "Google", "country": "USA", "date": "2019-04-12", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "First direct speech-to-speech translation without intermediate text", "source": "https://arxiv.org/abs/1904.06037" }, { "name": "OpenAI Five Finals", "org": "OpenAI", "country": "USA", "date": "2019-04-13", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "First AI to beat esports world champions (OG) at Dota 2", "source": "https://en.wikipedia.org/wiki/OpenAI_Five" }, { "name": "Sparse Transformer", "org": "OpenAI", "country": "USA", "date": "2019-04-23", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Sparse attention modeling far longer sequences of text, images, audio", "source": "https://openai.com/blog/sparse-transformer/" }, { "name": "MuseNet", "org": "OpenAI", "country": "USA", "date": "2019-04-25", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Generates four-minute multi-instrument compositions across styles", "source": "https://openai.com/index/musenet/" }, { "name": "ESM", "org": "Meta", "country": "USA", "date": "2019-04-29", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Protein language model on 250M sequences learns structure and function", "source": "https://www.biorxiv.org/content/10.1101/622803v1" }, { "name": "GPT-2 355M", "org": "OpenAI", "country": "USA", "date": "2019-05-01", "precision": "month", "category": "language", "open_weights": true, "params": "355M", "note": "Second stage of GPT-2 staged release (initially labeled 345M)", "source": "https://openai.com/blog/gpt-2-6-month-follow-up/" }, { "name": "SinGAN", "org": "Technion", "country": "Israel", "date": "2019-05-02", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Generative model learned from a single image; ICCV 2019 best paper", "source": "https://arxiv.org/abs/1905.01164" }, { "name": "MobileNetV3", "org": "Google", "country": "USA", "date": "2019-05-06", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "NAS-tuned efficient vision backbone for mobile phones", "source": "https://arxiv.org/abs/1905.02244" }, { "name": "UniLM", "org": "Microsoft", "country": "USA", "date": "2019-05-08", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Unified pretraining for language understanding and generation", "source": "https://arxiv.org/abs/1905.03197" }, { "name": "Few-Shot Talking Head Models", "org": "Samsung", "country": "South Korea", "date": "2019-05-20", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Animated realistic talking heads, even Mona Lisa, from few images", "source": "https://arxiv.org/abs/1905.08233" }, { "name": "FastSpeech", "org": "Microsoft", "country": "USA", "date": "2019-05-22", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Non-autoregressive transformer TTS enabling much faster speech synthesis", "source": "https://arxiv.org/abs/1905.09263" }, { "name": "EfficientNet", "org": "Google", "country": "USA", "date": "2019-05-28", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Compound scaling: state-of-the-art accuracy with far fewer parameters", "source": "https://arxiv.org/abs/1905.11946" }, { "name": "Grover", "org": "Allen Institute for AI", "country": "USA", "date": "2019-05-29", "precision": "day", "category": "language", "open_weights": true, "params": "1.5B", "note": "Neural fake-news generator doubling as its best detector", "source": "https://arxiv.org/abs/1905.12616" }, { "name": "VQ-VAE-2", "org": "DeepMind", "country": "UK", "date": "2019-06-02", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Hierarchical discrete latents rival GANs on image fidelity", "source": "https://arxiv.org/abs/1906.00446" }, { "name": "Chinese BERT-wwm", "org": "iFlytek", "country": "China", "date": "2019-06-19", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "HIT-iFLYTEK whole-word-masking Chinese BERT; widely used Chinese baseline", "source": "https://arxiv.org/abs/1906.08101" }, { "name": "XLNet", "org": "Google", "country": "USA", "date": "2019-06-19", "precision": "day", "category": "language", "open_weights": true, "params": "340M", "note": "Permutation language modeling (with CMU); beat BERT on 20 tasks", "source": "https://arxiv.org/abs/1906.08237" }, { "name": "Deep TabNine", "org": "TabNine", "country": "Canada", "date": "2019-07-01", "precision": "month", "category": "code", "open_weights": false, "params": "", "note": "GPT-2-based deep-learning code autocompletion across many languages", "source": "https://www.theverge.com/2019/7/24/20708542/coding-autocompleter-deep-tabnine-ai-deep-learning-smart-compose" }, { "name": "BigBiGAN", "org": "DeepMind", "country": "UK", "date": "2019-07-04", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "GAN-based unsupervised representation learning at scale", "source": "https://arxiv.org/abs/1907.02544" }, { "name": "Pluribus", "org": "Meta", "country": "USA", "date": "2019-07-11", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Facebook AI/CMU: first to beat pros at six-player no-limit hold'em", "source": "https://en.wikipedia.org/wiki/Pluribus_(poker_bot)" }, { "name": "NCSN", "org": "Stanford University", "country": "USA", "date": "2019-07-12", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Score-based generative modeling; key precursor to diffusion models", "source": "https://arxiv.org/abs/1907.05600" }, { "name": "DVD-GAN", "org": "DeepMind", "country": "UK", "date": "2019-07-15", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "GAN generating plausible high-resolution video on complex datasets", "source": "https://arxiv.org/abs/1907.06571" }, { "name": "SpanBERT", "org": "Meta", "country": "USA", "date": "2019-07-24", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Span-masking BERT variant excelling at question answering and coreference", "source": "https://arxiv.org/abs/1907.10529" }, { "name": "RoBERTa", "org": "Meta", "country": "USA", "date": "2019-07-26", "precision": "day", "category": "language", "open_weights": true, "params": "355M", "note": "Robustly optimized BERT pretraining; topped GLUE", "source": "https://arxiv.org/abs/1907.11692" }, { "name": "ERNIE 2.0", "org": "Baidu", "country": "China", "date": "2019-07-29", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Continual multi-task pretraining outperforming BERT and XLNet", "source": "https://arxiv.org/abs/1907.12412" }, { "name": "ViLBERT", "org": "Meta", "country": "USA", "date": "2019-08-06", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Two-stream BERT pretraining for joint vision-and-language tasks", "source": "https://arxiv.org/abs/1908.02265" }, { "name": "Megatron-LM", "org": "NVIDIA", "country": "USA", "date": "2019-08-13", "precision": "day", "category": "language", "open_weights": false, "params": "8.3B", "note": "Model-parallel training of an 8.3B GPT-2-style transformer", "source": "https://nv-adlr.github.io/MegatronLM" }, { "name": "StructBERT", "org": "Alibaba", "country": "China", "date": "2019-08-13", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Structure-aware BERT pretraining that topped the GLUE leaderboard", "source": "https://arxiv.org/abs/1908.04577" }, { "name": "GPT-2 774M", "org": "OpenAI", "country": "USA", "date": "2019-08-20", "precision": "day", "category": "language", "open_weights": true, "params": "774M", "note": "Third stage of GPT-2 staged release", "source": "https://en.wikipedia.org/wiki/GPT-2" }, { "name": "Sentence-BERT", "org": "TU Darmstadt (UKP Lab)", "country": "Germany", "date": "2019-08-27", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Siamese BERT sentence embeddings for fast semantic search", "source": "https://arxiv.org/abs/1908.10084" }, { "name": "DistilBERT", "org": "Hugging Face", "country": "USA", "date": "2019-08-28", "precision": "day", "category": "language", "open_weights": true, "params": "66M", "note": "Distilled BERT: 40% smaller, 60% faster, keeps 97% capability", "source": "https://medium.com/huggingface/distilbert-8cf3380435b5" }, { "name": "NEZHA", "org": "Huawei", "country": "China", "date": "2019-08-31", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Huawei Noah's Ark pretrained Chinese language model", "source": "https://arxiv.org/abs/1909.00204" }, { "name": "Aristo", "org": "Allen Institute for AI", "country": "USA", "date": "2019-09-04", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Scored over 90% on New York 8th-grade Regents science exam", "source": "https://arxiv.org/abs/1909.01958" }, { "name": "FermiNet", "org": "DeepMind", "country": "UK", "date": "2019-09-05", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Neural network solving the many-electron Schrodinger equation ab initio", "source": "https://arxiv.org/abs/1909.02487" }, { "name": "CTRL", "org": "Salesforce", "country": "USA", "date": "2019-09-11", "precision": "day", "category": "language", "open_weights": true, "params": "1.63B", "note": "Controllable generation via control codes; largest public LM then", "source": "https://arxiv.org/abs/1909.05858" }, { "name": "Hide-and-Seek agents", "org": "OpenAI", "country": "USA", "date": "2019-09-17", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Multi-agent RL agents spontaneously invented tool use", "source": "https://openai.com/blog/emergent-tool-use/" }, { "name": "TinyBERT", "org": "Huawei", "country": "China", "date": "2019-09-23", "precision": "day", "category": "language", "open_weights": true, "params": "14.5M", "note": "Distilled BERT 7.5x smaller and 9.4x faster at inference", "source": "https://arxiv.org/abs/1909.10351" }, { "name": "GAN-TTS", "org": "DeepMind", "country": "UK", "date": "2019-09-25", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Adversarially trained feed-forward speech synthesis rivaling WaveNet", "source": "https://arxiv.org/abs/1909.11646" }, { "name": "UNITER", "org": "Microsoft", "country": "USA", "date": "2019-09-25", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Universal image-text representation topping vision-language benchmarks", "source": "https://arxiv.org/abs/1909.11740" }, { "name": "ALBERT", "org": "Google", "country": "USA", "date": "2019-09-26", "precision": "day", "category": "language", "open_weights": true, "params": "235M", "note": "Parameter-sharing lite BERT topping GLUE, SQuAD and RACE", "source": "https://arxiv.org/abs/1909.11942" }, { "name": "MelGAN", "org": "Lyrebird AI", "country": "Canada", "date": "2019-10-08", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Fast non-autoregressive GAN vocoder for waveform synthesis", "source": "https://arxiv.org/abs/1910.06711" }, { "name": "Dactyl (Rubik's Cube)", "org": "OpenAI", "country": "USA", "date": "2019-10-15", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Robot hand solves Rubik's Cube via automatic domain randomization", "source": "https://openai.com/blog/solving-rubiks-cube/" }, { "name": "T5", "org": "Google", "country": "USA", "date": "2019-10-23", "precision": "day", "category": "language", "open_weights": true, "params": "11B", "note": "Unified text-to-text transformer; introduced the C4 dataset", "source": "https://arxiv.org/abs/1910.10683" }, { "name": "Parallel WaveGAN", "org": "LINE", "country": "Japan", "date": "2019-10-25", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Fast distillation-free GAN vocoder from LINE and NAVER", "source": "https://arxiv.org/abs/1910.11480" }, { "name": "Spleeter", "org": "Deezer", "country": "France", "date": "2019-10-28", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open-source pretrained music source separation into stems", "source": "https://pypi.org/project/spleeter/" }, { "name": "BART", "org": "Meta", "country": "USA", "date": "2019-10-29", "precision": "day", "category": "language", "open_weights": true, "params": "406M", "note": "Denoising seq2seq pretraining; strong at summarization and generation", "source": "https://arxiv.org/abs/1910.13461" }, { "name": "AlphaStar Final", "org": "DeepMind", "country": "UK", "date": "2019-10-30", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Reached Grandmaster level in StarCraft II on public ladder", "source": "https://www.deepmind.com/blog/alphastar-grandmaster-level-in-starcraft-ii-using-multi-agent-reinforcement-learning" }, { "name": "DialoGPT", "org": "Microsoft", "country": "USA", "date": "2019-11-01", "precision": "day", "category": "language", "open_weights": true, "params": "762M", "note": "GPT-2 trained on 147M Reddit exchanges for dialogue", "source": "https://arxiv.org/abs/1911.00536" }, { "name": "GPT-2 1.5B", "org": "OpenAI", "country": "USA", "date": "2019-11-05", "precision": "day", "category": "language", "open_weights": true, "params": "1.5B", "note": "Full GPT-2 weights released, completing the staged release", "source": "https://openai.com/blog/gpt-2-1-5b-release/" }, { "name": "XLM-R", "org": "Meta", "country": "USA", "date": "2019-11-05", "precision": "day", "category": "language", "open_weights": true, "params": "550M", "note": "Multilingual RoBERTa trained on 100 languages of CommonCrawl", "source": "https://arxiv.org/abs/1911.02116" }, { "name": "CamemBERT", "org": "Inria", "country": "France", "date": "2019-11-10", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "State-of-the-art French RoBERTa-style language model", "source": "https://arxiv.org/abs/1911.03894" }, { "name": "Noisy Student", "org": "Google", "country": "USA", "date": "2019-11-11", "precision": "day", "category": "vision", "open_weights": true, "params": "480M", "note": "Self-training EfficientNet-L2 set a new ImageNet state of the art", "source": "https://arxiv.org/abs/1911.04252" }, { "name": "Compressive Transformer", "org": "DeepMind", "country": "UK", "date": "2019-11-13", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Compressed long-range memory for book-length language modeling", "source": "https://arxiv.org/abs/1911.05507" }, { "name": "MoCo", "org": "Meta", "country": "USA", "date": "2019-11-13", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Momentum contrast: self-supervised visual pretraining rivaling supervised", "source": "https://arxiv.org/abs/1911.05722" }, { "name": "MuZero", "org": "DeepMind", "country": "UK", "date": "2019-11-19", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Masters Go, chess, shogi and Atari without knowing the rules", "source": "https://arxiv.org/abs/1911.08265" }, { "name": "EfficientDet", "org": "Google", "country": "USA", "date": "2019-11-20", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Scalable, efficient object detectors built on EfficientNet", "source": "https://arxiv.org/abs/1911.09070" }, { "name": "Dreamer", "org": "Google", "country": "USA", "date": "2019-12-03", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Agent learns control behaviors by imagining in a learned world model", "source": "https://arxiv.org/abs/1912.01603" }, { "name": "StyleGAN2", "org": "NVIDIA", "country": "USA", "date": "2019-12-03", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Fixed StyleGAN artifacts; new state of the art for faces", "source": "https://arxiv.org/abs/1912.04958" }, { "name": "PPLM", "org": "Uber", "country": "USA", "date": "2019-12-04", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Plug-and-play attribute models steering GPT-2 without retraining", "source": "https://arxiv.org/abs/1912.02164" }, { "name": "PEGASUS", "org": "Google", "country": "USA", "date": "2019-12-18", "precision": "day", "category": "language", "open_weights": true, "params": "568M", "note": "Gap-sentence pretraining for state-of-the-art abstractive summarization", "source": "https://arxiv.org/abs/1912.08777" }, { "name": "Juewu (Honor of Kings AI)", "org": "Tencent", "country": "China", "date": "2019-12-20", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Deep RL agent beating top pros at 1v1 Honor of Kings", "source": "https://arxiv.org/abs/1912.09729" }, { "name": "BiT (Big Transfer)", "org": "Google", "country": "USA", "date": "2019-12-24", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Large-scale supervised pretraining for strong few-shot visual transfer", "source": "https://arxiv.org/abs/1912.11370" }, { "name": "LayoutLM", "org": "Microsoft", "country": "USA", "date": "2019-12-31", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Joint text-and-layout pretraining for document image understanding", "source": "https://arxiv.org/abs/1912.13318" }, { "name": "Reformer", "org": "Google", "country": "USA", "date": "2020-01-13", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Efficient transformer using LSH attention and reversible layers", "source": "https://arxiv.org/abs/2001.04451" }, { "name": "DDSP", "org": "Google", "country": "USA", "date": "2020-01-14", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Magenta differentiable DSP for realistic neural audio synthesis", "source": "https://arxiv.org/abs/2001.04643" }, { "name": "mBART", "org": "Meta", "country": "USA", "date": "2020-01-22", "precision": "day", "category": "language", "open_weights": true, "params": "680M", "note": "Multilingual denoising seq2seq pretraining for machine translation", "source": "https://arxiv.org/abs/2001.08210" }, { "name": "Meena", "org": "Google", "country": "USA", "date": "2020-01-27", "precision": "day", "category": "language", "open_weights": false, "params": "2.6B", "note": "Open-domain chatbot; introduced the SSA human-likeness metric", "source": "https://arxiv.org/abs/2001.09977" }, { "name": "REALM", "org": "Google", "country": "USA", "date": "2020-02-10", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Retrieval-augmented language model pretraining for open-domain QA", "source": "https://arxiv.org/abs/2002.08909" }, { "name": "Turing-NLG", "org": "Microsoft", "country": "USA", "date": "2020-02-10", "precision": "day", "category": "language", "open_weights": false, "params": "17B", "note": "Largest language model at announcement, trained with DeepSpeed ZeRO", "source": "https://venturebeat.com/ai/microsoft-trains-worlds-largest-transformer-language-model/" }, { "name": "SimCLR", "org": "Google", "country": "USA", "date": "2020-02-13", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Simple contrastive self-supervised learning matching supervised baselines", "source": "https://arxiv.org/abs/2002.05709" }, { "name": "CodeBERT", "org": "Microsoft", "country": "USA", "date": "2020-02-19", "precision": "day", "category": "code", "open_weights": true, "params": "125M", "note": "Bimodal pretrained model for programming and natural languages", "source": "https://arxiv.org/abs/2002.08155" }, { "name": "MiniLM", "org": "Microsoft", "country": "USA", "date": "2020-02-25", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Deep self-attention distillation; compact models behind popular embeddings", "source": "https://arxiv.org/abs/2002.10957" }, { "name": "ELECTRA", "org": "Google", "country": "USA", "date": "2020-03-10", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Replaced-token detection pretraining, far more compute-efficient than BERT", "source": "https://research.google/blog/more-efficient-nlp-model-pre-training-with-electra/" }, { "name": "ProGen", "org": "Salesforce", "country": "USA", "date": "2020-03-13", "precision": "day", "category": "science", "open_weights": false, "params": "1.2B", "note": "Language model generating functional proteins", "source": "https://www.biorxiv.org/content/10.1101/2020.03.07.982272v2" }, { "name": "NeRF", "org": "UC Berkeley", "country": "USA", "date": "2020-03-19", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Neural radiance fields for photorealistic novel-view synthesis", "source": "https://arxiv.org/abs/2003.08934" }, { "name": "DLSS 2.0", "org": "NVIDIA", "country": "USA", "date": "2020-03-23", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Generalized AI upscaler for games, no per-game training needed", "source": "https://www.nvidia.com/en-us/geforce/news/nvidia-dlss-2-0-a-big-leap-in-ai-rendering/" }, { "name": "MetNet", "org": "Google", "country": "USA", "date": "2020-03-24", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Neural precipitation model beating physics forecasts up to 8 hours", "source": "https://arxiv.org/abs/2003.12140" }, { "name": "Agent57", "org": "DeepMind", "country": "UK", "date": "2020-03-30", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "First agent above human baseline on all 57 Atari games", "source": "https://deepmind.google/blog/agent57-outperforming-the-human-atari-benchmark/" }, { "name": "Suphx", "org": "Microsoft", "country": "USA", "date": "2020-03-30", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Microsoft Research Asia Mahjong AI reaching 10 dan on Tenhou", "source": "https://arxiv.org/abs/2003.13590" }, { "name": "Dense Passage Retrieval (DPR)", "org": "Meta", "country": "USA", "date": "2020-04-10", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Dual-encoder dense retriever outperforming BM25 for open-domain QA", "source": "https://arxiv.org/abs/2004.04906" }, { "name": "Longformer", "org": "Allen Institute for AI", "country": "USA", "date": "2020-04-10", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Sparse-attention transformer for long documents", "source": "https://arxiv.org/abs/2004.05150" }, { "name": "AlphaChip", "org": "Google", "country": "USA", "date": "2020-04-22", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "RL chip floorplanning used in TPU designs; named AlphaChip in 2024", "source": "https://arxiv.org/abs/2004.10746" }, { "name": "YOLOv4", "org": "Academia Sinica", "country": "Taiwan", "date": "2020-04-23", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Optimal speed-accuracy real-time object detector", "source": "https://arxiv.org/abs/2004.10934" }, { "name": "BlenderBot", "org": "Meta", "country": "USA", "date": "2020-04-28", "precision": "day", "category": "language", "open_weights": true, "params": "9.4B", "note": "Largest open-domain chatbot released, blending conversational skills", "source": "https://arxiv.org/abs/2004.13637" }, { "name": "Jukebox", "org": "OpenAI", "country": "USA", "date": "2020-04-30", "precision": "day", "category": "audio", "open_weights": true, "params": "5B", "note": "Generates raw-audio music with singing across genres and artists", "source": "https://openai.com/blog/jukebox/" }, { "name": "Conformer", "org": "Google", "country": "USA", "date": "2020-05-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Convolution-augmented transformer; state-of-the-art speech recognition", "source": "https://arxiv.org/abs/2005.08100" }, { "name": "GameGAN", "org": "NVIDIA", "country": "USA", "date": "2020-05-22", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Neural world model recreated Pac-Man purely from watching gameplay", "source": "https://blogs.nvidia.com/blog/2020/05/22/gamegan-research-pacman-anniversary/" }, { "name": "RAG", "org": "Meta", "country": "USA", "date": "2020-05-22", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Retrieval-augmented generation combining parametric and retrieved memory", "source": "https://arxiv.org/abs/2005.11401" }, { "name": "DETR", "org": "Meta", "country": "USA", "date": "2020-05-26", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "End-to-end transformer object detection without hand-designed anchors", "source": "https://arxiv.org/abs/2005.12872" }, { "name": "YOLOv5", "org": "Ultralytics", "country": "USA", "date": "2020-05-27", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Popular PyTorch real-time detector; repo public May 27, v1.0 June 25", "source": "https://github.com/ultralytics/yolov5/releases/tag/v1.0" }, { "name": "GPT-3", "org": "OpenAI", "country": "USA", "date": "2020-05-28", "precision": "day", "category": "language", "open_weights": false, "params": "175B", "note": "Few-shot in-context learning at 175B; API launched June 11", "source": "https://arxiv.org/abs/2005.14165" }, { "name": "DeBERTa", "org": "Microsoft", "country": "USA", "date": "2020-06-05", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Disentangled attention; later first single model to surpass humans on SuperGLUE", "source": "https://arxiv.org/abs/2006.03654" }, { "name": "TransCoder", "org": "Meta", "country": "USA", "date": "2020-06-05", "precision": "day", "category": "code", "open_weights": true, "params": "", "note": "Unsupervised translation between C++, Java and Python", "source": "https://arxiv.org/abs/2006.03511" }, { "name": "FastSpeech 2", "org": "Microsoft", "country": "USA", "date": "2020-06-08", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Fast non-autoregressive TTS with pitch and energy conditioning, widely adopted", "source": "https://arxiv.org/abs/2006.04558" }, { "name": "StyleGAN2-ADA", "org": "NVIDIA", "country": "USA", "date": "2020-06-11", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Adaptive augmentation trains GANs from a few thousand images", "source": "https://arxiv.org/abs/2006.06676" }, { "name": "BYOL", "org": "DeepMind", "country": "UK", "date": "2020-06-13", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Self-supervised image representations without negative pairs", "source": "https://arxiv.org/abs/2006.07733" }, { "name": "Image GPT", "org": "OpenAI", "country": "USA", "date": "2020-06-17", "precision": "day", "category": "image", "open_weights": true, "params": "6.8B", "note": "GPT-2-style transformer on raw pixels learns image representations", "source": "https://openai.com/blog/image-gpt/" }, { "name": "SimCLRv2", "org": "Google", "country": "USA", "date": "2020-06-17", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Big self-supervised models make strong semi-supervised learners", "source": "https://arxiv.org/abs/2006.10029" }, { "name": "DDPM", "org": "UC Berkeley", "country": "USA", "date": "2020-06-19", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Denoising diffusion reaches GAN-level quality; launched the diffusion era", "source": "https://arxiv.org/abs/2006.11239" }, { "name": "wav2vec 2.0", "org": "Meta", "country": "USA", "date": "2020-06-20", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Self-supervised speech; strong ASR from minutes of labeled data", "source": "https://arxiv.org/abs/2006.11477" }, { "name": "GShard", "org": "Google", "country": "USA", "date": "2020-06-30", "precision": "day", "category": "language", "open_weights": false, "params": "600B", "note": "600B sparse mixture-of-experts translation model across 100 languages", "source": "https://arxiv.org/abs/2006.16668" }, { "name": "PLATO-2", "org": "Baidu", "country": "China", "date": "2020-06-30", "precision": "day", "category": "language", "open_weights": true, "params": "1.6B", "note": "Chinese and English open-domain chatbot trained via curriculum learning", "source": "https://arxiv.org/abs/2006.16779" }, { "name": "ProtTrans", "org": "Technical University of Munich", "country": "Germany", "date": "2020-07-13", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Protein language models trained on supercomputers, released openly", "source": "https://arxiv.org/abs/2007.06225" }, { "name": "ReBeL", "org": "Meta", "country": "USA", "date": "2020-07-27", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "RL plus search for imperfect-information games; superhuman heads-up poker", "source": "https://arxiv.org/abs/2007.13544" }, { "name": "BigBird", "org": "Google", "country": "USA", "date": "2020-07-28", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Sparse-attention transformer handling much longer sequences", "source": "https://arxiv.org/abs/2007.14062" }, { "name": "GPT-f", "org": "OpenAI", "country": "USA", "date": "2020-09-07", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Transformer theorem prover; proofs accepted into Metamath library", "source": "https://arxiv.org/abs/2009.03393" }, { "name": "DiffWave", "org": "NVIDIA", "country": "USA", "date": "2020-09-21", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Diffusion model for audio waveforms matching WaveNet vocoder quality", "source": "https://arxiv.org/abs/2009.09761" }, { "name": "Performer", "org": "Google", "country": "USA", "date": "2020-09-30", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "FAVOR+ linear-complexity attention approximating softmax transformers", "source": "https://arxiv.org/abs/2009.14794" }, { "name": "DreamerV2", "org": "Google", "country": "USA", "date": "2020-10-05", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "World-model agent reaching human-level Atari performance", "source": "https://arxiv.org/abs/2010.02193" }, { "name": "HiFi-GAN", "org": "Kakao Enterprise", "country": "South Korea", "date": "2020-10-12", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Efficient high-fidelity GAN vocoder widely adopted in TTS", "source": "https://arxiv.org/abs/2010.05646" }, { "name": "M2M-100", "org": "Meta", "country": "USA", "date": "2020-10-19", "precision": "day", "category": "language", "open_weights": true, "params": "15B", "note": "Many-to-many translation across 100 languages without English pivot", "source": "https://ai.meta.com/blog/introducing-many-to-many-multilingual-machine-translation/" }, { "name": "mT5", "org": "Google", "country": "USA", "date": "2020-10-22", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Multilingual T5 covering 101 languages", "source": "https://arxiv.org/abs/2010.11934" }, { "name": "ruGPT-3", "org": "Sber", "country": "Russia", "date": "2020-10-22", "precision": "day", "category": "language", "open_weights": true, "params": "760M", "note": "Sber's open Russian GPT-3-style models (Large released publicly)", "source": "https://habr.com/ru/companies/sberdevices/articles/524522/" }, { "name": "Vision Transformer (ViT)", "org": "Google", "country": "USA", "date": "2020-10-22", "precision": "day", "category": "vision", "open_weights": true, "params": "632M", "note": "Pure transformer on image patches rivals CNNs at scale", "source": "https://arxiv.org/abs/2010.11929" }, { "name": "AlphaFold 2", "org": "DeepMind", "country": "UK", "date": "2020-11-30", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Solved protein structure prediction at CASP14; open-sourced July 2021", "source": "https://deepmind.google/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology/" }, { "name": "CPM", "org": "Tsinghua University", "country": "China", "date": "2020-12-01", "precision": "day", "category": "language", "open_weights": true, "params": "2.6B", "note": "Tsinghua/BAAI: largest Chinese generative pretrained LM at release", "source": "https://arxiv.org/abs/2012.00413" }, { "name": "ESM-1b", "org": "Meta", "country": "USA", "date": "2020-12-01", "precision": "month", "category": "science", "open_weights": true, "params": "650M", "note": "Widely used open protein language model from Facebook AI", "source": "https://github.com/facebookresearch/esm" }, { "name": "VQGAN", "org": "Heidelberg University", "country": "Germany", "date": "2020-12-17", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Taming Transformers: codebook GAN plus transformer for high-res images", "source": "https://arxiv.org/abs/2012.09841" }, { "name": "DeiT", "org": "Meta", "country": "USA", "date": "2020-12-23", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Data-efficient vision transformers trained on ImageNet alone", "source": "https://arxiv.org/abs/2012.12877" }, { "name": "CLIP", "org": "OpenAI", "country": "USA", "date": "2021-01-05", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Contrastive image-text pretraining enabling zero-shot image classification", "source": "https://openai.com/blog/clip/" }, { "name": "DALL·E", "org": "OpenAI", "country": "USA", "date": "2021-01-05", "precision": "day", "category": "image", "open_weights": false, "params": "12B", "note": "Landmark text-to-image generation from a 12B GPT-3 variant", "source": "https://openai.com/blog/dall-e/" }, { "name": "Switch Transformer", "org": "Google", "country": "USA", "date": "2021-01-11", "precision": "day", "category": "language", "open_weights": true, "params": "1.6T", "note": "First trillion-parameter sparse mixture-of-experts language model", "source": "https://arxiv.org/abs/2101.03961" }, { "name": "Wu Dao 1.0", "org": "BAAI", "country": "China", "date": "2021-01-11", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "BAAI super-scale model suite; first announced Jan 2021, full suite March", "source": "https://en.wikipedia.org/wiki/Wu_Dao" }, { "name": "TimeSformer", "org": "Meta", "country": "USA", "date": "2021-02-09", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Convolution-free video transformer with divided space-time attention", "source": "https://arxiv.org/abs/2102.05095" }, { "name": "ALIGN", "org": "Google", "country": "USA", "date": "2021-02-11", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Image-text model trained on 1.8B noisy alt-text pairs", "source": "https://arxiv.org/abs/2102.05918" }, { "name": "NFNets", "org": "DeepMind", "country": "UK", "date": "2021-02-11", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Normalizer-free ResNets set ImageNet state of the art without batch norm", "source": "https://arxiv.org/abs/2102.06171" }, { "name": "Improved DDPM", "org": "OpenAI", "country": "USA", "date": "2021-02-18", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Learned variances enable faster, higher-likelihood diffusion sampling", "source": "https://arxiv.org/abs/2102.09672" }, { "name": "M6", "org": "Alibaba", "country": "China", "date": "2021-03-01", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Large Chinese multimodal pretrainer for text and images", "source": "https://arxiv.org/abs/2103.00823" }, { "name": "SEER", "org": "Meta", "country": "USA", "date": "2021-03-02", "precision": "day", "category": "vision", "open_weights": true, "params": "1.3B", "note": "Self-supervised on a billion uncurated Instagram images", "source": "https://arxiv.org/abs/2103.01988" }, { "name": "Perceiver", "org": "DeepMind", "country": "UK", "date": "2021-03-04", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "One architecture for images, audio, video and point clouds", "source": "https://arxiv.org/abs/2103.03206" }, { "name": "GLM", "org": "Tsinghua University", "country": "China", "date": "2021-03-18", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Autoregressive blank-infilling pretraining; foundation of later ChatGLM", "source": "https://arxiv.org/abs/2103.10360" }, { "name": "GPT-Neo", "org": "EleutherAI", "country": "USA", "date": "2021-03-21", "precision": "day", "category": "language", "open_weights": true, "params": "2.7B", "note": "First open-source GPT-3-style models, trained on the Pile", "source": "https://github.com/EleutherAI/gpt-neo" }, { "name": "Swin Transformer", "org": "Microsoft", "country": "USA", "date": "2021-03-25", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Hierarchical shifted-window vision transformer; ICCV 2021 best paper", "source": "https://arxiv.org/abs/2103.14030" }, { "name": "EfficientNetV2", "org": "Google", "country": "USA", "date": "2021-04-01", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Smaller, faster-training ConvNets via training-aware architecture search", "source": "https://arxiv.org/abs/2104.00298" }, { "name": "DGMR", "org": "DeepMind", "country": "UK", "date": "2021-04-02", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Generative radar rain nowcasting preferred by Met Office forecasters", "source": "https://arxiv.org/abs/2104.00954" }, { "name": "SR3", "org": "Google", "country": "USA", "date": "2021-04-15", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Diffusion-based image super-resolution beating GANs in human evaluation", "source": "https://arxiv.org/abs/2104.07636" }, { "name": "MT-Opt", "org": "Google", "country": "USA", "date": "2021-04-16", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Multi-task robotic RL across a fleet of real robots", "source": "https://arxiv.org/abs/2104.08212" }, { "name": "VideoGPT", "org": "UC Berkeley", "country": "USA", "date": "2021-04-20", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "VQ-VAE plus transformer baseline for video generation", "source": "https://arxiv.org/abs/2104.10157" }, { "name": "PanGu-α", "org": "Huawei", "country": "China", "date": "2021-04-26", "precision": "day", "category": "language", "open_weights": true, "params": "200B", "note": "200B Chinese LM trained on Ascend chips; smaller checkpoints released", "source": "https://arxiv.org/abs/2104.12369" }, { "name": "DINO", "org": "Meta", "country": "USA", "date": "2021-04-29", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Self-distilled ViTs learn segmentation-like attention without labels", "source": "https://arxiv.org/abs/2104.14294" }, { "name": "GODIVA", "org": "Microsoft", "country": "USA", "date": "2021-04-30", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Early open-domain text-to-video generation model", "source": "https://arxiv.org/abs/2104.14806" }, { "name": "MLP-Mixer", "org": "Google", "country": "USA", "date": "2021-05-04", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "All-MLP vision architecture competitive with CNNs and ViTs", "source": "https://arxiv.org/abs/2105.01601" }, { "name": "ADM (Guided Diffusion)", "org": "OpenAI", "country": "USA", "date": "2021-05-11", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Diffusion models beat GANs on ImageNet image synthesis", "source": "https://arxiv.org/abs/2105.05233" }, { "name": "LaMDA", "org": "Google", "country": "USA", "date": "2021-05-18", "precision": "day", "category": "language", "open_weights": false, "params": "137B", "note": "Dialogue-focused LLM unveiled at Google I/O 2021", "source": "https://blog.google/technology/ai/lamda/" }, { "name": "MUM", "org": "Google", "country": "USA", "date": "2021-05-18", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Multitask Unified Model for Search: multilingual and multimodal", "source": "https://blog.google/products/search/introducing-mum/" }, { "name": "wav2vec-U", "org": "Meta", "country": "USA", "date": "2021-05-21", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Speech recognition trained without any transcribed audio", "source": "https://ai.meta.com/blog/wav2vec-unsupervised-speech-recognition-without-supervision/" }, { "name": "HyperCLOVA", "org": "Naver", "country": "South Korea", "date": "2021-05-25", "precision": "day", "category": "language", "open_weights": false, "params": "204B", "note": "Korean GPT-3-class model trained largely on Korean data", "source": "https://en.yna.co.kr/view/AEN20210525005400320" }, { "name": "CogView", "org": "Tsinghua University", "country": "China", "date": "2021-05-26", "precision": "day", "category": "image", "open_weights": true, "params": "4B", "note": "4B-parameter Chinese text-to-image transformer", "source": "https://arxiv.org/abs/2105.13290" }, { "name": "ByT5", "org": "Google", "country": "USA", "date": "2021-05-28", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Token-free byte-level T5 models", "source": "https://arxiv.org/abs/2105.13626" }, { "name": "Wu Dao 2.0", "org": "BAAI", "country": "China", "date": "2021-06-01", "precision": "day", "category": "multimodal", "open_weights": false, "params": "1.75T", "note": "1.75-trillion-parameter multimodal MoE, largest model announced then", "source": "https://www.engadget.com/chinas-gigantic-multi-modal-ai-is-no-one-trick-pony-211414388.html" }, { "name": "Decision Transformer", "org": "UC Berkeley", "country": "USA", "date": "2021-06-02", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Recasts reinforcement learning as return-conditioned sequence modeling", "source": "https://arxiv.org/abs/2106.01345" }, { "name": "GPT-J", "org": "EleutherAI", "country": "USA", "date": "2021-06-04", "precision": "day", "category": "language", "open_weights": true, "params": "6B", "note": "Strongest open GPT-3-style model of 2021, trained in JAX on TPUs", "source": "https://arankomatsuzaki.wordpress.com/2021/06/04/gpt-j/" }, { "name": "ViT-G/14", "org": "Google", "country": "USA", "date": "2021-06-08", "precision": "day", "category": "vision", "open_weights": false, "params": "2B", "note": "Scaling Vision Transformers: 2B ViT hits 90.45% ImageNet top-1", "source": "https://arxiv.org/abs/2106.04560" }, { "name": "CoAtNet", "org": "Google", "country": "USA", "date": "2021-06-09", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Convolution-attention hybrid reaching 90.88% ImageNet top-1", "source": "https://arxiv.org/abs/2106.04803" }, { "name": "VITS", "org": "Kakao Enterprise", "country": "South Korea", "date": "2021-06-11", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "End-to-end TTS producing natural, humanlike speech", "source": "https://arxiv.org/abs/2106.06103" }, { "name": "HuBERT", "org": "Meta", "country": "USA", "date": "2021-06-14", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Masked-prediction self-supervised speech representations", "source": "https://arxiv.org/abs/2106.07447" }, { "name": "BEiT", "org": "Microsoft", "country": "USA", "date": "2021-06-15", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "BERT-style masked image modeling for vision transformers", "source": "https://arxiv.org/abs/2106.08254" }, { "name": "RoseTTAFold", "org": "University of Washington", "country": "USA", "date": "2021-06-15", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open three-track protein structure network nearing AlphaFold 2 accuracy", "source": "https://www.biorxiv.org/content/10.1101/2021.06.14.448402v1" }, { "name": "CPM-2", "org": "Tsinghua University", "country": "China", "date": "2021-06-20", "precision": "day", "category": "language", "open_weights": true, "params": "11B", "note": "Cost-effective Chinese-English models, including a 198B MoE variant", "source": "https://arxiv.org/abs/2106.10715" }, { "name": "StyleGAN3", "org": "NVIDIA", "country": "USA", "date": "2021-06-23", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Alias-free GAN for animation-ready image synthesis", "source": "https://arxiv.org/abs/2106.12423" }, { "name": "Frozen", "org": "DeepMind", "country": "UK", "date": "2021-06-25", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Vision encoder prompts a frozen LM for few-shot multimodal learning", "source": "https://arxiv.org/abs/2106.13884" }, { "name": "GitHub Copilot", "org": "GitHub", "country": "USA", "date": "2021-06-29", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "AI pair programmer preview; debuted OpenAI Codex publicly", "source": "https://github.blog/2021-06-29-introducing-github-copilot-ai-pair-programmer/" }, { "name": "DALL·E Mini", "org": "Craiyon", "country": "USA", "date": "2021-07-01", "precision": "month", "category": "image", "open_weights": true, "params": "", "note": "Open DALL·E replica built at HF JAX/Flax week; viral in 2022", "source": "https://github.com/borisdayma/dalle-mini" }, { "name": "ERNIE 3.0", "org": "Baidu", "country": "China", "date": "2021-07-05", "precision": "day", "category": "language", "open_weights": false, "params": "10B", "note": "Knowledge-enhanced 10B model that topped SuperGLUE", "source": "https://arxiv.org/abs/2107.02137" }, { "name": "Codex", "org": "OpenAI", "country": "USA", "date": "2021-07-07", "precision": "day", "category": "code", "open_weights": false, "params": "12B", "note": "GPT fine-tuned on GitHub code; introduced HumanEval; API August 10", "source": "https://arxiv.org/abs/2107.03374" }, { "name": "SoundStream", "org": "Google", "country": "USA", "date": "2021-07-07", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "End-to-end neural audio codec underpinning later audio language models", "source": "https://arxiv.org/abs/2107.03312" }, { "name": "BlenderBot 2.0", "org": "Meta", "country": "USA", "date": "2021-07-16", "precision": "day", "category": "language", "open_weights": true, "params": "2.7B", "note": "Open chatbot with long-term memory and live internet search", "source": "https://ai.meta.com/blog/blender-bot-2-an-open-source-chatbot-that-builds-long-term-memory-and-searches-the-internet/" }, { "name": "YOLOX", "org": "Megvii", "country": "China", "date": "2021-07-18", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Anchor-free YOLO detector outperforming YOLOv5", "source": "https://arxiv.org/abs/2107.08430" }, { "name": "XLand agents", "org": "DeepMind", "country": "UK", "date": "2021-07-27", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Generally capable agents from open-ended play across procedural games", "source": "https://deepmind.google/blog/generally-capable-agents-emerge-from-open-ended-play/" }, { "name": "Perceiver IO", "org": "DeepMind", "country": "UK", "date": "2021-07-30", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Perceiver generalized to arbitrary structured outputs across modalities", "source": "https://arxiv.org/abs/2107.14795" }, { "name": "Jurassic-1", "org": "AI21 Labs", "country": "Israel", "date": "2021-08-11", "precision": "day", "category": "language", "open_weights": false, "params": "178B", "note": "178B Jumbo model launched with AI21 Studio as GPT-3 rival", "source": "https://en.wikipedia.org/wiki/AI21_Labs" }, { "name": "SimVLM", "org": "Google", "country": "USA", "date": "2021-08-24", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Prefix-LM vision-language pretraining with weak supervision; zero-shot VQA", "source": "https://arxiv.org/abs/2108.10904" }, { "name": "CodeT5", "org": "Salesforce", "country": "USA", "date": "2021-09-02", "precision": "day", "category": "code", "open_weights": true, "params": "220M", "note": "Identifier-aware encoder-decoder for code understanding and generation", "source": "https://arxiv.org/abs/2109.00859" }, { "name": "FLAN", "org": "Google", "country": "USA", "date": "2021-09-03", "precision": "day", "category": "language", "open_weights": false, "params": "137B", "note": "Instruction tuning boosts zero-shot performance of a 137B LM", "source": "https://arxiv.org/abs/2109.01652" }, { "name": "Macaw", "org": "Allen Institute for AI", "country": "USA", "date": "2021-09-06", "precision": "day", "category": "language", "open_weights": true, "params": "11B", "note": "Open general-purpose QA model outperforming GPT-3 on some probes", "source": "https://arxiv.org/abs/2109.02593" }, { "name": "PLATO-XL", "org": "Baidu", "country": "China", "date": "2021-09-20", "precision": "day", "category": "language", "open_weights": false, "params": "11B", "note": "11B Chinese-English dialogue model, largest chatbot model at release", "source": "https://arxiv.org/abs/2109.09519" }, { "name": "AlphaFold-Multimer", "org": "DeepMind", "country": "UK", "date": "2021-10-04", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "AlphaFold extended to predicting protein complexes", "source": "https://www.biorxiv.org/content/10.1101/2021.10.04.463034v1" }, { "name": "Enformer", "org": "DeepMind", "country": "UK", "date": "2021-10-04", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Transformer predicting gene expression from long DNA sequences", "source": "https://deepmind.google/blog/predicting-gene-expression-with-ai/" }, { "name": "M6-10T", "org": "Alibaba", "country": "China", "date": "2021-10-08", "precision": "day", "category": "multimodal", "open_weights": false, "params": "10T", "note": "Ten-trillion-parameter multimodal MoE pretrained on 512 GPUs", "source": "https://arxiv.org/abs/2110.03888" }, { "name": "Yuan 1.0", "org": "Inspur", "country": "China", "date": "2021-10-10", "precision": "day", "category": "language", "open_weights": false, "params": "245B", "note": "245B Chinese language model, largest Chinese LM at the time", "source": "https://arxiv.org/abs/2110.04725" }, { "name": "Megatron-Turing NLG 530B", "org": "Microsoft", "country": "USA", "date": "2021-10-11", "precision": "day", "category": "language", "open_weights": false, "params": "530B", "note": "Microsoft and NVIDIA's 530B model, largest dense LM then", "source": "https://www.microsoft.com/en-us/research/blog/using-deepspeed-and-megatron-to-train-megatron-turing-nlg-530b-the-worlds-largest-and-most-powerful-generative-language-model/" }, { "name": "T0", "org": "Hugging Face", "country": "USA", "date": "2021-10-15", "precision": "day", "category": "language", "open_weights": true, "params": "11B", "note": "BigScience multitask-prompted T5 generalizing zero-shot to unseen tasks", "source": "https://arxiv.org/abs/2110.08207" }, { "name": "WavLM", "org": "Microsoft", "country": "USA", "date": "2021-10-26", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Self-supervised speech model topping SUPERB across full-stack tasks", "source": "https://arxiv.org/abs/2110.13900" }, { "name": "EfficientZero", "org": "Tsinghua University", "country": "China", "date": "2021-10-30", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Superhuman Atari performance from two hours of gameplay", "source": "https://arxiv.org/abs/2111.00210" }, { "name": "ruDALL-E", "org": "Sber", "country": "Russia", "date": "2021-11-02", "precision": "day", "category": "image", "open_weights": true, "params": "1.3B", "note": "Open Russian text-to-image model (Malevich)", "source": "https://github.com/ai-forever/ru-dalle" }, { "name": "MAE", "org": "Meta", "country": "USA", "date": "2021-11-11", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Masked autoencoders: scalable self-supervised vision pretraining", "source": "https://arxiv.org/abs/2111.06377" }, { "name": "KoGPT", "org": "Kakao Brain", "country": "South Korea", "date": "2021-11-12", "precision": "day", "category": "language", "open_weights": true, "params": "6B", "note": "Open 6B Korean GPT-3-style language model", "source": "https://github.com/kakaobrain/kogpt" }, { "name": "MetNet-2", "org": "Google", "country": "USA", "date": "2021-11-14", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Twelve-hour precipitation forecasts outperforming physics-based models", "source": "https://arxiv.org/abs/2111.07470" }, { "name": "LiT", "org": "Google", "country": "USA", "date": "2021-11-15", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Locked-image text tuning; 85.2% zero-shot ImageNet accuracy", "source": "https://arxiv.org/abs/2111.07991" }, { "name": "INTERN", "org": "Shanghai AI Lab", "country": "China", "date": "2021-11-16", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "General vision foundation model with SenseTime and partners", "source": "https://arxiv.org/abs/2111.08687" }, { "name": "XLS-R", "org": "Meta", "country": "USA", "date": "2021-11-17", "precision": "day", "category": "audio", "open_weights": true, "params": "300M-2B", "note": "Cross-lingual speech model covering 128 languages", "source": "https://arxiv.org/abs/2111.09296" }, { "name": "DeBERTaV3", "org": "Microsoft", "country": "USA", "date": "2021-11-18", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "ELECTRA-style pretraining with gradient-disentangled embedding sharing", "source": "https://arxiv.org/abs/2111.09543" }, { "name": "Swin Transformer V2", "org": "Microsoft", "country": "USA", "date": "2021-11-18", "precision": "day", "category": "vision", "open_weights": true, "params": "3B", "note": "3B-parameter SwinV2-G, largest dense vision model at the time", "source": "https://arxiv.org/abs/2111.09883" }, { "name": "Florence", "org": "Microsoft", "country": "USA", "date": "2021-11-22", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Computer vision foundation model spanning images, video and text", "source": "https://arxiv.org/abs/2111.11432" }, { "name": "GauGAN2", "org": "NVIDIA", "country": "USA", "date": "2021-11-22", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Public demo generating landscapes from text and sketches", "source": "https://blogs.nvidia.com/blog/2021/11/22/gaugan2-ai-art-demo/" }, { "name": "NÜWA", "org": "Microsoft", "country": "USA", "date": "2021-11-24", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Unified model for text-to-image and text-to-video generation", "source": "https://arxiv.org/abs/2111.12417" }, { "name": "EXAONE", "org": "LG AI Research", "country": "South Korea", "date": "2021-12-01", "precision": "month", "category": "multimodal", "open_weights": false, "params": "", "note": "LG's first expert AI foundation model, unveiled December 2021", "source": "https://ko.wikipedia.org/wiki/엑사원" }, { "name": "minDALL-E", "org": "Kakao Brain", "country": "South Korea", "date": "2021-12-01", "precision": "month", "category": "image", "open_weights": true, "params": "1.3B", "note": "Open 1.3B text-to-image model trained on 14M image-text pairs", "source": "https://github.com/kakaobrain/minDALL-E" }, { "name": "Player of Games", "org": "DeepMind", "country": "UK", "date": "2021-12-06", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "One algorithm mastering perfect- and imperfect-information games", "source": "https://arxiv.org/abs/2112.03178" }, { "name": "Gopher", "org": "DeepMind", "country": "UK", "date": "2021-12-08", "precision": "day", "category": "language", "open_weights": false, "params": "280B", "note": "DeepMind's 280B LM with extensive scaling and ethics analysis", "source": "https://deepmind.google/blog/language-modelling-at-scale-gopher-ethical-considerations-and-retrieval/" }, { "name": "RETRO", "org": "DeepMind", "country": "UK", "date": "2021-12-08", "precision": "day", "category": "language", "open_weights": false, "params": "7.5B", "note": "Retrieval from 2T-token database rivals models 25x larger", "source": "https://arxiv.org/abs/2112.04426" }, { "name": "DM21", "org": "DeepMind", "country": "UK", "date": "2021-12-09", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Neural density functional improving quantum chemistry simulations", "source": "https://deepmind.google/blog/simulating-matter-on-the-quantum-scale-with-ai/" }, { "name": "GLaM", "org": "Google", "country": "USA", "date": "2021-12-09", "precision": "day", "category": "language", "open_weights": false, "params": "1.2T (97B active)", "note": "Sparse MoE beating GPT-3 with a third of training energy", "source": "https://research.google/blog/more-efficient-in-context-learning-with-glam/" }, { "name": "WebGPT", "org": "OpenAI", "country": "USA", "date": "2021-12-16", "precision": "day", "category": "agent", "open_weights": false, "params": "175B", "note": "GPT-3 fine-tuned to browse the web and cite sources", "source": "https://openai.com/blog/webgpt/" }, { "name": "GLIDE", "org": "OpenAI", "country": "USA", "date": "2021-12-20", "precision": "day", "category": "image", "open_weights": true, "params": "3.5B", "note": "Text-guided diffusion for photorealistic generation and editing", "source": "https://arxiv.org/abs/2112.10741" }, { "name": "Latent Diffusion Models", "org": "LMU Munich", "country": "Germany", "date": "2021-12-20", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "CompVis diffusion in latent space; direct precursor of Stable Diffusion", "source": "https://arxiv.org/abs/2112.10752" }, { "name": "XGLM", "org": "Meta", "country": "USA", "date": "2021-12-20", "precision": "day", "category": "language", "open_weights": true, "params": "7.5B", "note": "Open multilingual GPT-style models; few-shot SOTA in 20+ languages", "source": "https://arxiv.org/abs/2112.10668" }, { "name": "ERNIE 3.0 Titan", "org": "Baidu", "country": "China", "date": "2021-12-23", "precision": "day", "category": "language", "open_weights": false, "params": "260B", "note": "260B knowledge-enhanced Chinese LM built with Peng Cheng Lab", "source": "https://arxiv.org/abs/2112.12731" }, { "name": "ERNIE-ViLG", "org": "Baidu", "country": "China", "date": "2021-12-31", "precision": "day", "category": "image", "open_weights": false, "params": "10B", "note": "10B Chinese bidirectional text-to-image and image-to-text model", "source": "https://arxiv.org/abs/2112.15283" }, { "name": "Detic", "org": "Meta", "country": "USA", "date": "2022-01-07", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Open-vocabulary detector recognizing 20,000+ classes via image-level supervision", "source": "https://arxiv.org/abs/2201.02605" }, { "name": "ConvNeXt", "org": "Meta", "country": "USA", "date": "2022-01-10", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Modernized pure ConvNet matching Vision Transformers on ImageNet", "source": "https://arxiv.org/abs/2201.03545" }, { "name": "Instant NGP", "org": "NVIDIA", "country": "USA", "date": "2022-01-16", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Hash-encoded neural graphics primitives training NeRFs in seconds (Instant NeRF)", "source": "https://arxiv.org/abs/2201.05989" }, { "name": "CM3", "org": "Meta", "country": "USA", "date": "2022-01-19", "precision": "day", "category": "multimodal", "open_weights": false, "params": "13B", "note": "Causally masked model generating text and images from web HTML", "source": "https://arxiv.org/abs/2201.07520" }, { "name": "data2vec", "org": "Meta", "country": "USA", "date": "2022-01-20", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "One self-supervised learning method for speech, vision and text", "source": "https://ai.facebook.com/research/data2vec-a-general-framework-for-self-supervised-learning-in-speech-vision-and-language/" }, { "name": "cpt-text / cpt-code (OpenAI Embeddings)", "org": "OpenAI", "country": "USA", "date": "2022-01-24", "precision": "day", "category": "language", "open_weights": false, "params": "300M-175B", "note": "First OpenAI embeddings API models for text and code search", "source": "https://arxiv.org/abs/2201.10005" }, { "name": "InstructGPT", "org": "OpenAI", "country": "USA", "date": "2022-01-27", "precision": "day", "category": "language", "open_weights": false, "params": "1.3B-175B", "note": "RLHF-tuned GPT-3 made API default; direct precursor to ChatGPT", "source": "https://en.wikipedia.org/wiki/2022_in_artificial_intelligence" }, { "name": "BLIP", "org": "Salesforce", "country": "USA", "date": "2022-01-28", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Bootstrapped vision-language pretraining for captioning and retrieval", "source": "https://arxiv.org/abs/2201.12086" }, { "name": "Midjourney V1", "org": "Midjourney", "country": "USA", "date": "2022-02-01", "precision": "month", "category": "image", "open_weights": false, "params": "", "note": "First Midjourney image model, in closed Discord beta", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "AlphaCode", "org": "DeepMind", "country": "UK", "date": "2022-02-02", "precision": "day", "category": "code", "open_weights": false, "params": "41B", "note": "Competitive programming at roughly median human Codeforces level", "source": "https://deepmind.google/discover/blog/competitive-programming-with-alphacode/" }, { "name": "GPT-NeoX-20B", "org": "EleutherAI", "country": "USA", "date": "2022-02-02", "precision": "day", "category": "language", "open_weights": true, "params": "20B", "note": "Largest open-weights GPT-style model at release; weights out Feb 9", "source": "https://blog.eleuther.ai/announcing-20b/" }, { "name": "GPT-f (Lean, statement curriculum learning)", "org": "OpenAI", "country": "USA", "date": "2022-02-03", "precision": "day", "category": "science", "open_weights": false, "params": "774M", "note": "Expert iteration solved several formal math olympiad problems in Lean", "source": "https://arxiv.org/abs/2202.01344" }, { "name": "M3GNet", "org": "UC San Diego", "country": "USA", "date": "2022-02-05", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Universal graph interatomic potential spanning the periodic table", "source": "https://arxiv.org/abs/2202.02450" }, { "name": "OFA", "org": "Alibaba", "country": "China", "date": "2022-02-07", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Unified seq2seq model across vision, language and multimodal tasks", "source": "https://arxiv.org/abs/2202.03052" }, { "name": "MaskGIT", "org": "Google", "country": "USA", "date": "2022-02-08", "precision": "day", "category": "image", "open_weights": true, "params": "227M", "note": "Masked generative image transformer with parallel decoding", "source": "https://arxiv.org/abs/2202.04200" }, { "name": "GT Sophy", "org": "Sony AI", "country": "Japan", "date": "2022-02-09", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Deep RL agent beat champion Gran Turismo drivers; Nature paper", "source": "https://doi.org/10.1038/s41586-021-04357-7" }, { "name": "MuZero-RC", "org": "DeepMind", "country": "UK", "date": "2022-02-11", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "MuZero's first real-world use: YouTube VP9 video compression rate control", "source": "https://deepmind.google/discover/blog/muzeros-first-step-from-research-into-the-real-world/" }, { "name": "SEER 10B", "org": "Meta", "country": "USA", "date": "2022-02-16", "precision": "day", "category": "vision", "open_weights": true, "params": "10B", "note": "Self-supervised vision model trained on uncurated internet images", "source": "https://arxiv.org/abs/2202.08360" }, { "name": "TCV tokamak plasma controller (deep RL)", "org": "DeepMind", "country": "UK", "date": "2022-02-16", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Deep RL controller shaped plasma in EPFL's TCV fusion tokamak", "source": "https://deepmind.google/blog/accelerating-fusion-science-through-learned-plasma-control/" }, { "name": "ST-MoE", "org": "Google", "country": "USA", "date": "2022-02-17", "precision": "day", "category": "language", "open_weights": false, "params": "269B", "note": "Stable, transferable sparse mixture-of-experts; SuperGLUE state of the art", "source": "https://arxiv.org/abs/2202.08906" }, { "name": "FourCastNet", "org": "NVIDIA", "country": "USA", "date": "2022-02-22", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Fourier neural operator global weather model, far faster than NWP", "source": "https://arxiv.org/abs/2202.11214" }, { "name": "PolyCoder", "org": "Carnegie Mellon University", "country": "USA", "date": "2022-02-26", "precision": "day", "category": "code", "open_weights": true, "params": "2.7B", "note": "Open code model over 12 languages; beat Codex on C", "source": "https://arxiv.org/abs/2202.13169" }, { "name": "Ithaca", "org": "DeepMind", "country": "UK", "date": "2022-03-09", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Restores, dates and locates ancient Greek inscriptions; Nature paper", "source": "https://deepmind.google/blog/predicting-the-past-with-ithaca/" }, { "name": "ProtGPT2", "org": "University of Bayreuth", "country": "Germany", "date": "2022-03-12", "precision": "day", "category": "science", "open_weights": true, "params": "738M", "note": "GPT-2-style protein language model generating novel sequences", "source": "https://www.biorxiv.org/content/10.1101/2022.03.09.483666v1" }, { "name": "code-davinci-002", "org": "OpenAI", "country": "USA", "date": "2022-03-15", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Codex model later described as the GPT-3.5 base model", "source": "https://en.wikipedia.org/wiki/GPT-3" }, { "name": "text-davinci-002", "org": "OpenAI", "country": "USA", "date": "2022-03-15", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "First GPT-3.5 series model, launched with edit and insert modes", "source": "https://en.wikipedia.org/wiki/GPT-3" }, { "name": "GopherCite", "org": "DeepMind", "country": "UK", "date": "2022-03-16", "precision": "day", "category": "language", "open_weights": false, "params": "280B", "note": "Gopher trained with RLHF to back answers with verified quotes", "source": "https://deepmind.google/discover/blog/gophercite-teaching-language-models-to-support-answers-with-verified-quotes/" }, { "name": "Z-code MoE", "org": "Microsoft", "country": "USA", "date": "2022-03-22", "precision": "day", "category": "language", "open_weights": false, "params": "5B", "note": "Mixture-of-experts translation models deployed in production Microsoft Translator", "source": "https://www.microsoft.com/en-us/research/blog/microsoft-translator-enhanced-with-z-code-mixture-of-experts-models/" }, { "name": "Make-A-Scene", "org": "Meta", "country": "USA", "date": "2022-03-24", "precision": "day", "category": "image", "open_weights": false, "params": "4B", "note": "Text-to-image generation controllable by scene sketches", "source": "https://arxiv.org/abs/2203.13131" }, { "name": "CodeGen", "org": "Salesforce", "country": "USA", "date": "2022-03-25", "precision": "day", "category": "code", "open_weights": true, "params": "16.1B", "note": "Open code LLM family for multi-turn program synthesis", "source": "https://arxiv.org/abs/2203.13474" }, { "name": "Chinchilla", "org": "DeepMind", "country": "UK", "date": "2022-03-29", "precision": "day", "category": "language", "open_weights": false, "params": "70B", "note": "Compute-optimal scaling: 70B model beat 280B Gopher", "source": "https://arxiv.org/abs/2203.15556" }, { "name": "Tortoise TTS", "org": "James Betker", "country": "USA", "date": "2022-04-01", "precision": "month", "category": "audio", "open_weights": true, "params": "", "note": "Influential open multi-voice TTS cloning voices from a few clips", "source": "https://nonint.com/2022/04/25/tortoise-architectural-design-doc/" }, { "name": "PaLM", "org": "Google", "country": "USA", "date": "2022-04-04", "precision": "day", "category": "language", "open_weights": false, "params": "540B", "note": "540B dense model trained with Pathways on TPU v4 pods", "source": "https://arxiv.org/abs/2204.02311" }, { "name": "SayCan", "org": "Google", "country": "USA", "date": "2022-04-04", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Grounds LLM task plans in robot affordances", "source": "https://arxiv.org/abs/2204.01691" }, { "name": "DALL·E 2", "org": "OpenAI", "country": "USA", "date": "2022-04-06", "precision": "day", "category": "image", "open_weights": false, "params": "3.5B", "note": "Diffusion text-to-image; beta July, open to all September 2022", "source": "https://en.wikipedia.org/wiki/DALL-E" }, { "name": "Video Diffusion Models", "org": "Google", "country": "USA", "date": "2022-04-07", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Early demonstration of diffusion-based text-to-video generation", "source": "https://arxiv.org/abs/2204.03458" }, { "name": "InCoder", "org": "Meta", "country": "USA", "date": "2022-04-12", "precision": "day", "category": "code", "open_weights": true, "params": "6.7B", "note": "Open code model supporting infilling as well as left-to-right generation", "source": "https://arxiv.org/abs/2204.05999" }, { "name": "Midjourney V2", "org": "Midjourney", "country": "USA", "date": "2022-04-12", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Second Midjourney model, more refined and colorful", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "VLM-4", "org": "LightOn", "country": "France", "date": "2022-04-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "LLMs trained natively in five European languages, via Muse API", "source": "https://lighton.ai/blog/lighton-publicly-launches-muse-an-api-to-vlm-4-large-language-models-trained-natively-in-five-european-languages/" }, { "name": "NOOR", "org": "TII", "country": "UAE", "date": "2022-04-13", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Billed as the world's largest Arabic NLP model at launch", "source": "https://en.wikipedia.org/wiki/Technology_Innovation_Institute" }, { "name": "mGPT", "org": "Sber", "country": "Russia", "date": "2022-04-15", "precision": "day", "category": "language", "open_weights": true, "params": "1.3B-13B", "note": "Open multilingual GPT covering 60 languages from 25 families", "source": "https://arxiv.org/abs/2204.07580" }, { "name": "CogView2", "org": "Tsinghua University", "country": "China", "date": "2022-04-28", "precision": "day", "category": "image", "open_weights": true, "params": "6B", "note": "Hierarchical-transformer Chinese/English text-to-image model", "source": "https://arxiv.org/abs/2204.14217" }, { "name": "Flamingo", "org": "DeepMind", "country": "UK", "date": "2022-04-28", "precision": "day", "category": "multimodal", "open_weights": false, "params": "80B", "note": "Few-shot visual language model over interleaved images and text", "source": "https://deepmind.google/discover/blog/tackling-multiple-tasks-with-a-single-visual-language-model/" }, { "name": "Jurassic-X", "org": "AI21 Labs", "country": "Israel", "date": "2022-05-01", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "MRKL neuro-symbolic system routing LLM queries to tools", "source": "https://arxiv.org/abs/2205.00445" }, { "name": "OPT", "org": "Meta", "country": "USA", "date": "2022-05-03", "precision": "day", "category": "language", "open_weights": true, "params": "125M-175B", "note": "Open GPT-3-scale replication; OPT-175B weights shared with researchers", "source": "https://ai.meta.com/blog/democratizing-access-to-large-scale-language-models-with-opt-175b/" }, { "name": "CoCa", "org": "Google", "country": "USA", "date": "2022-05-04", "precision": "day", "category": "vision", "open_weights": false, "params": "2.1B", "note": "Contrastive captioner image-text foundation model; 91% ImageNet top-1", "source": "https://arxiv.org/abs/2205.01917" }, { "name": "NaturalSpeech", "org": "Microsoft", "country": "USA", "date": "2022-05-09", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "End-to-end TTS reaching human-level quality on LJSpeech", "source": "https://arxiv.org/abs/2205.04421" }, { "name": "UL2", "org": "Google", "country": "USA", "date": "2022-05-10", "precision": "day", "category": "language", "open_weights": true, "params": "20B", "note": "Mixture-of-denoisers pretraining; 20B checkpoint open-sourced", "source": "https://arxiv.org/abs/2205.05131" }, { "name": "LaMDA 2", "org": "Google", "country": "USA", "date": "2022-05-11", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Unveiled at Google I/O alongside the AI Test Kitchen app", "source": "https://en.wikipedia.org/wiki/LaMDA" }, { "name": "Gato", "org": "DeepMind", "country": "UK", "date": "2022-05-12", "precision": "day", "category": "multimodal", "open_weights": false, "params": "1.2B", "note": "Single generalist agent across 604 tasks: games, chat, robotics", "source": "https://deepmind.google/blog/a-generalist-agent/" }, { "name": "OWL-ViT", "org": "Google", "country": "USA", "date": "2022-05-12", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Simple open-vocabulary object detection with vision transformers", "source": "https://arxiv.org/abs/2205.06230" }, { "name": "HyperTree Proof Search (HTPS)", "org": "Meta", "country": "USA", "date": "2022-05-23", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Online-trained neural theorem prover for Lean and Metamath", "source": "https://arxiv.org/abs/2205.11491" }, { "name": "Imagen", "org": "Google", "country": "USA", "date": "2022-05-23", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Photorealistic text-to-image diffusion using frozen T5-XXL text encoder", "source": "https://arxiv.org/abs/2205.11487" }, { "name": "GIT", "org": "Microsoft", "country": "USA", "date": "2022-05-27", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Generative image-to-text transformer for captioning and VQA", "source": "https://arxiv.org/abs/2205.14100" }, { "name": "CogVideo", "org": "Tsinghua University", "country": "China", "date": "2022-05-29", "precision": "day", "category": "video", "open_weights": true, "params": "9.4B", "note": "Large open text-to-video transformer built on CogView2", "source": "https://arxiv.org/abs/2205.15868" }, { "name": "DALL·E Mega", "org": "Craiyon", "country": "USA", "date": "2022-06-01", "precision": "month", "category": "image", "open_weights": true, "params": "", "note": "Larger DALL·E Mini successor behind the viral Craiyon generator", "source": "https://huggingface.co/dalle-mini/dalle-mega" }, { "name": "GPT-SW3", "org": "AI Sweden", "country": "Sweden", "date": "2022-06-01", "precision": "month", "category": "language", "open_weights": false, "params": "3.5B", "note": "First large generative language model for Swedish", "source": "https://aclanthology.org/2022.lrec-1.376/" }, { "name": "ProteinMPNN", "org": "University of Washington", "country": "USA", "date": "2022-06-04", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Baker lab deep learning protein sequence design", "source": "https://doi.org/10.1101/2022.06.03.494563" }, { "name": "MineDojo (MineCLIP)", "org": "NVIDIA", "country": "USA", "date": "2022-06-17", "precision": "day", "category": "games", "open_weights": true, "params": "", "note": "Minecraft agent framework with internet-scale MineCLIP reward model", "source": "https://arxiv.org/abs/2206.08853" }, { "name": "Unified-IO", "org": "Allen Institute for AI", "country": "USA", "date": "2022-06-17", "precision": "day", "category": "multimodal", "open_weights": true, "params": "2.9B", "note": "One seq2seq model across vision, language and multimodal tasks", "source": "https://arxiv.org/abs/2206.08916" }, { "name": "GODEL", "org": "Microsoft", "country": "USA", "date": "2022-06-22", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Open grounded dialogue pretrained models for goal-directed chat", "source": "https://arxiv.org/abs/2206.11309" }, { "name": "OpenFold", "org": "Columbia University", "country": "USA", "date": "2022-06-22", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Trainable open-source AlphaFold2 reproduction released with original weights", "source": "https://github.com/aqlaboratory/openfold/releases/tag/v1.0.0" }, { "name": "Parti", "org": "Google", "country": "USA", "date": "2022-06-22", "precision": "day", "category": "image", "open_weights": false, "params": "20B", "note": "Autoregressive 20B-parameter text-to-image model", "source": "https://arxiv.org/abs/2206.10789" }, { "name": "Amazon CodeWhisperer", "org": "Amazon", "country": "USA", "date": "2022-06-23", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "AWS ML coding companion debuting Amazon's code model, in preview", "source": "https://aws.amazon.com/about-aws/whats-new/2022/06/aws-announces-amazon-codewhisperer-preview/" }, { "name": "Video PreTraining (VPT)", "org": "OpenAI", "country": "USA", "date": "2022-06-23", "precision": "day", "category": "games", "open_weights": true, "params": "", "note": "Learned Minecraft from unlabeled video; crafted diamond tools", "source": "https://arxiv.org/abs/2206.11795" }, { "name": "YaLM 100B", "org": "Yandex", "country": "Russia", "date": "2022-06-23", "precision": "day", "category": "language", "open_weights": true, "params": "100B", "note": "Large open GPT-like model for Russian and English", "source": "https://medium.com/yandex/yandex-publishes-yalm-100b-its-the-largest-gpt-like-neural-network-in-open-source-d1df53d0e9a6" }, { "name": "YOLOv6", "org": "Meituan", "country": "China", "date": "2022-06-23", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Industrial real-time object detector released by Meituan", "source": "https://github.com/meituan/YOLOv6/releases" }, { "name": "ProGen2", "org": "Salesforce", "country": "USA", "date": "2022-06-27", "precision": "day", "category": "science", "open_weights": true, "params": "6.4B", "note": "Protein language models scaled to 6.4B parameters", "source": "https://arxiv.org/abs/2206.13517" }, { "name": "Minerva", "org": "Google", "country": "USA", "date": "2022-06-29", "precision": "day", "category": "science", "open_weights": false, "params": "540B", "note": "PaLM tuned on math and science papers; 50% on MATH", "source": "https://arxiv.org/abs/2206.14858" }, { "name": "DeepNash", "org": "DeepMind", "country": "UK", "date": "2022-06-30", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Model-free RL reached human expert level at Stratego", "source": "https://arxiv.org/abs/2206.15378" }, { "name": "CodeRL", "org": "Salesforce", "country": "USA", "date": "2022-07-05", "precision": "day", "category": "code", "open_weights": true, "params": "770M", "note": "Code generation trained with RL from unit-test feedback", "source": "https://arxiv.org/abs/2207.01780" }, { "name": "NLLB-200", "org": "Meta", "country": "USA", "date": "2022-07-06", "precision": "day", "category": "language", "open_weights": true, "params": "54.5B", "note": "Open machine translation model covering 200 languages", "source": "https://research.facebook.com/publications/no-language-left-behind/" }, { "name": "YOLOv7", "org": "Academia Sinica", "country": "Taiwan", "date": "2022-07-06", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Real-time object detector setting speed-accuracy state of the art", "source": "https://arxiv.org/abs/2207.02696" }, { "name": "BLOOM", "org": "Hugging Face", "country": "USA", "date": "2022-07-12", "precision": "day", "category": "language", "open_weights": true, "params": "176B", "note": "BigScience open multilingual model built by 1,000+ researchers", "source": "https://huggingface.co/blog/bloom" }, { "name": "Inner Monologue", "org": "Google", "country": "USA", "date": "2022-07-12", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "LLM robot planning with closed-loop language feedback", "source": "https://arxiv.org/abs/2207.05608" }, { "name": "NUWA-Infinity", "org": "Microsoft", "country": "USA", "date": "2022-07-20", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Autoregressive-over-autoregressive synthesis of arbitrarily large images and video", "source": "https://arxiv.org/abs/2207.09814" }, { "name": "ESM-2 / ESMFold", "org": "Meta", "country": "USA", "date": "2022-07-21", "precision": "day", "category": "science", "open_weights": true, "params": "15B", "note": "Protein language model enabling fast single-sequence structure prediction", "source": "https://doi.org/10.1101/2022.07.20.500902" }, { "name": "OmegaFold", "org": "Helixon", "country": "China", "date": "2022-07-22", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Single-sequence protein structure prediction without alignments", "source": "https://doi.org/10.1101/2022.07.21.500999" }, { "name": "PanGu-Coder", "org": "Huawei", "country": "China", "date": "2022-07-22", "precision": "day", "category": "code", "open_weights": false, "params": "2.6B", "note": "Huawei function-level Python code generation model", "source": "https://arxiv.org/abs/2207.11280" }, { "name": "Midjourney V3", "org": "Midjourney", "country": "USA", "date": "2022-07-25", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Default model during Midjourney's open beta surge", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "GPT-NeoX-Japanese", "org": "ABEJA", "country": "Japan", "date": "2022-07-27", "precision": "day", "category": "language", "open_weights": true, "params": "2.7B", "note": "Japanese GPT-NeoX models from ABEJA; 2.7B later open-sourced", "source": "https://tech-blog.abeja.asia/entry/abeja-gpt-project-202207" }, { "name": "HelixFold-Single", "org": "Baidu", "country": "China", "date": "2022-07-28", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "MSA-free protein structure prediction via protein language model", "source": "https://arxiv.org/abs/2207.13921" }, { "name": "AlexaTM 20B", "org": "Amazon", "country": "USA", "date": "2022-08-02", "precision": "day", "category": "language", "open_weights": false, "params": "20B", "note": "Multilingual seq2seq model beating PaLM 540B on 1-shot summarization", "source": "https://arxiv.org/abs/2208.01448" }, { "name": "Prompt-to-Prompt", "org": "Google", "country": "USA", "date": "2022-08-02", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Text-driven image editing via cross-attention control", "source": "https://arxiv.org/abs/2208.01626" }, { "name": "Textual Inversion", "org": "Tel Aviv University", "country": "Israel", "date": "2022-08-02", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Personalizes text-to-image models by learning new word embeddings", "source": "https://arxiv.org/abs/2208.01618" }, { "name": "GLM-130B", "org": "Zhipu AI", "country": "China", "date": "2022-08-04", "precision": "day", "category": "language", "open_weights": true, "params": "130B", "note": "Open bilingual Chinese-English 130B model from Tsinghua and Zhipu", "source": "https://keg.cs.tsinghua.edu.cn/glm-130b/posts/glm-130b/" }, { "name": "Atlas", "org": "Meta", "country": "USA", "date": "2022-08-05", "precision": "day", "category": "language", "open_weights": true, "params": "11B", "note": "Retrieval-augmented 11B model beating PaLM 540B on few-shot QA", "source": "https://arxiv.org/abs/2208.03299" }, { "name": "BlenderBot 3", "org": "Meta", "country": "USA", "date": "2022-08-05", "precision": "day", "category": "language", "open_weights": true, "params": "175B", "note": "Public chatbot demo with internet search and continual learning", "source": "https://arxiv.org/abs/2208.03188" }, { "name": "Stable Diffusion", "org": "Stability AI", "country": "UK", "date": "2022-08-10", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open latent diffusion text-to-image; public v1.4 weights followed Aug 22", "source": "https://stability.ai/news-updates/stable-diffusion-announcement" }, { "name": "Luminous-supreme", "org": "Aleph Alpha", "country": "Germany", "date": "2022-08-15", "precision": "month", "category": "language", "open_weights": false, "params": "70B", "note": "Largest model in Aleph Alpha's European multilingual Luminous family", "source": "https://web.archive.org/web/20220818084714/https://www.aleph-alpha.com/pricing" }, { "name": "PaLM-SayCan", "org": "Google", "country": "USA", "date": "2022-08-16", "precision": "day", "category": "robotics", "open_weights": false, "params": "540B", "note": "SayCan robot planner upgraded with PaLM for grounded instructions", "source": "https://arxiv.org/abs/2204.01691" }, { "name": "BEiT-3", "org": "Microsoft", "country": "USA", "date": "2022-08-22", "precision": "day", "category": "multimodal", "open_weights": true, "params": "1.9B", "note": "Multiway transformer; state of the art on vision and vision-language", "source": "https://arxiv.org/abs/2208.10442" }, { "name": "DreamBooth", "org": "Google", "country": "USA", "date": "2022-08-25", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Subject-driven fine-tuning of text-to-image diffusion models", "source": "https://arxiv.org/abs/2208.12242" }, { "name": "AudioLM", "org": "Google", "country": "USA", "date": "2022-09-07", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Language-model approach to coherent speech and piano continuation", "source": "https://arxiv.org/abs/2209.03143" }, { "name": "Japanese Stable Diffusion", "org": "rinna", "country": "Japan", "date": "2022-09-09", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Stable Diffusion adapted for Japanese-language prompts", "source": "https://huggingface.co/rinna/japanese-stable-diffusion" }, { "name": "ACT-1", "org": "Adept", "country": "USA", "date": "2022-09-14", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Action Transformer operating web software from natural language", "source": "https://www.adept.ai/act" }, { "name": "PaLI", "org": "Google", "country": "USA", "date": "2022-09-14", "precision": "day", "category": "multimodal", "open_weights": false, "params": "17B", "note": "Jointly scaled multilingual language-image model", "source": "https://arxiv.org/abs/2209.06794" }, { "name": "NeMo Megatron GPT-20B", "org": "NVIDIA", "country": "USA", "date": "2022-09-15", "precision": "day", "category": "language", "open_weights": true, "params": "20B", "note": "NVIDIA open 20B GPT checkpoint for the NeMo framework", "source": "https://huggingface.co/nvidia/nemo-megatron-gpt-20B" }, { "name": "OpenCLIP ViT-H/14", "org": "LAION", "country": "Germany", "date": "2022-09-15", "precision": "day", "category": "vision", "open_weights": true, "params": "986M", "note": "Largest open CLIP at release; text encoder for Stable Diffusion 2", "source": "https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K" }, { "name": "Character.AI (beta)", "org": "Character.AI", "country": "USA", "date": "2022-09-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Persona chatbot on in-house LLM from ex-LaMDA developers; public beta", "source": "https://en.wikipedia.org/wiki/Character.ai" }, { "name": "Code as Policies", "org": "Google", "country": "USA", "date": "2022-09-16", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "LLM writes robot policy code from natural language", "source": "https://arxiv.org/abs/2209.07753" }, { "name": "CodeGeeX", "org": "Zhipu AI", "country": "China", "date": "2022-09-19", "precision": "day", "category": "code", "open_weights": true, "params": "13B", "note": "Multilingual code model trained on Huawei Ascend chips, with Tsinghua", "source": "https://models.aminer.cn/codegeex/blog/" }, { "name": "WeLM", "org": "Tencent", "country": "China", "date": "2022-09-21", "precision": "day", "category": "language", "open_weights": false, "params": "10B", "note": "WeChat AI Chinese LLM rivaling models up to 25x larger", "source": "https://arxiv.org/abs/2209.10372" }, { "name": "Whisper", "org": "OpenAI", "country": "USA", "date": "2022-09-21", "precision": "day", "category": "audio", "open_weights": true, "params": "1.55B", "note": "Open multilingual speech recognition trained on 680k hours", "source": "https://en.wikipedia.org/wiki/Whisper_(speech_recognition_system)" }, { "name": "CPM-Ant", "org": "OpenBMB", "country": "China", "date": "2022-09-22", "precision": "day", "category": "language", "open_weights": true, "params": "10B", "note": "Chinese 10B model trained live in public (CPM-Live)", "source": "https://github.com/OpenBMB/CPM-Live/tree/cpm-ant/cpm-live#model-checkpoints" }, { "name": "GET3D", "org": "NVIDIA", "country": "USA", "date": "2022-09-22", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Generates textured 3D meshes learned from 2D images", "source": "https://arxiv.org/abs/2209.11163" }, { "name": "Sparrow", "org": "DeepMind", "country": "UK", "date": "2022-09-22", "precision": "day", "category": "language", "open_weights": false, "params": "70B", "note": "Chinchilla-based dialogue agent trained with RLHF and rules", "source": "https://deepmind.google/discover/blog/building-safer-dialogue-agents/" }, { "name": "Dramatron", "org": "DeepMind", "country": "UK", "date": "2022-09-29", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Hierarchical LLM tool for co-writing screenplays and theatre scripts", "source": "https://arxiv.org/abs/2209.14958" }, { "name": "DreamFusion", "org": "Google", "country": "USA", "date": "2022-09-29", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Text-to-3D via score distillation from Imagen", "source": "https://arxiv.org/abs/2209.14988" }, { "name": "Make-A-Video", "org": "Meta", "country": "USA", "date": "2022-09-29", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Text-to-video built without paired text-video data", "source": "https://arxiv.org/abs/2209.14792" }, { "name": "AudioGen", "org": "Meta", "country": "USA", "date": "2022-09-30", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Autoregressive text-to-sound-effects generator", "source": "https://arxiv.org/abs/2209.15352" }, { "name": "NovelAI Diffusion", "org": "NovelAI", "country": "USA", "date": "2022-10-03", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Anime-focused image model; weights leaked days after launch", "source": "https://en.wikipedia.org/wiki/NovelAI" }, { "name": "DiffDock", "org": "MIT", "country": "USA", "date": "2022-10-04", "precision": "day", "category": "science", "open_weights": true, "params": "20M", "note": "Diffusion generative model for molecular docking", "source": "https://arxiv.org/abs/2210.01776" }, { "name": "AlphaTensor", "org": "DeepMind", "country": "UK", "date": "2022-10-05", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "RL discovered faster matrix multiplication algorithms; Nature paper", "source": "https://www.nature.com/articles/s41586-022-05172-4" }, { "name": "Imagen Video", "org": "Google", "country": "USA", "date": "2022-10-05", "precision": "day", "category": "video", "open_weights": false, "params": "11.6B", "note": "Cascaded diffusion generating high-definition video from text", "source": "https://arxiv.org/abs/2210.02303" }, { "name": "Phenaki", "org": "Google", "country": "USA", "date": "2022-10-05", "precision": "day", "category": "video", "open_weights": false, "params": "1.8B", "note": "Variable-length video from sequences of text prompts", "source": "https://arxiv.org/abs/2210.02399" }, { "name": "VIMA", "org": "NVIDIA", "country": "USA", "date": "2022-10-06", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Robot manipulation from multimodal prompts, with Stanford", "source": "https://arxiv.org/abs/2210.03094" }, { "name": "GenSLM", "org": "Argonne National Laboratory", "country": "USA", "date": "2022-10-11", "precision": "day", "category": "science", "open_weights": true, "params": "25B", "note": "Genome-scale language models tracking SARS-CoV-2 evolution", "source": "https://doi.org/10.1101/2022.10.10.511571" }, { "name": "Imagic", "org": "Google", "country": "USA", "date": "2022-10-17", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Text-based editing of real photos with diffusion models", "source": "https://arxiv.org/abs/2210.09276" }, { "name": "BioGPT", "org": "Microsoft", "country": "USA", "date": "2022-10-19", "precision": "day", "category": "language", "open_weights": true, "params": "347M", "note": "GPT pretrained on PubMed for biomedical text generation and mining", "source": "https://arxiv.org/abs/2210.10341" }, { "name": "Hokkien speech-to-speech translation", "org": "Meta", "country": "USA", "date": "2022-10-19", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "First AI speech translation system for a primarily oral language", "source": "https://ai.meta.com/blog/ai-translation-hokkien/" }, { "name": "Flan-PaLM", "org": "Google", "country": "USA", "date": "2022-10-20", "precision": "day", "category": "language", "open_weights": false, "params": "540B", "note": "PaLM instruction-tuned on 1,800+ tasks with chain-of-thought data", "source": "https://arxiv.org/abs/2210.11416" }, { "name": "Flan-T5", "org": "Google", "country": "USA", "date": "2022-10-20", "precision": "day", "category": "language", "open_weights": true, "params": "80M-11B", "note": "Open instruction-tuned T5 checkpoints, widely used baseline", "source": "https://arxiv.org/abs/2210.11416" }, { "name": "Stable Diffusion v1.5", "org": "Runway", "country": "USA", "date": "2022-10-20", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Most widely fine-tuned Stable Diffusion 1.x checkpoint", "source": "https://en.wikipedia.org/wiki/Stable_Diffusion" }, { "name": "U-PaLM", "org": "Google", "country": "USA", "date": "2022-10-20", "precision": "day", "category": "language", "open_weights": false, "params": "540B", "note": "PaLM continued with UL2 objective for ~0.1% extra compute", "source": "https://arxiv.org/abs/2210.11399" }, { "name": "EnCodec", "org": "Meta", "country": "USA", "date": "2022-10-24", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Neural audio codec later used as tokenizer for audio LMs", "source": "https://arxiv.org/abs/2210.13438" }, { "name": "ERNIE-ViLG 2.0", "org": "Baidu", "country": "China", "date": "2022-10-27", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Knowledge-enhanced Chinese text-to-image diffusion with expert denoisers", "source": "https://arxiv.org/abs/2210.15257" }, { "name": "Taiyi Stable Diffusion", "org": "IDEA", "country": "China", "date": "2022-10-31", "precision": "day", "category": "image", "open_weights": true, "params": "1B", "note": "Chinese-language Stable Diffusion from the Fengshenbang project", "source": "https://huggingface.co/IDEA-CCNL/Taiyi-Stable-Diffusion-1B-Chinese-v0.1" }, { "name": "AltDiffusion", "org": "BAAI", "country": "China", "date": "2022-11-01", "precision": "month", "category": "image", "open_weights": true, "params": "", "note": "Multilingual Stable Diffusion variant built on AltCLIP", "source": "https://huggingface.co/BAAI/AltDiffusion" }, { "name": "Chinese CLIP", "org": "Alibaba", "country": "China", "date": "2022-11-02", "precision": "day", "category": "vision", "open_weights": true, "params": "77M-958M", "note": "Open CLIP models pretrained on Chinese image-text pairs", "source": "https://arxiv.org/abs/2211.01335" }, { "name": "eDiff-I", "org": "NVIDIA", "country": "USA", "date": "2022-11-02", "precision": "day", "category": "image", "open_weights": false, "params": "9.1B", "note": "Text-to-image with an ensemble of expert denoisers", "source": "https://arxiv.org/abs/2211.01324" }, { "name": "BLOOMZ", "org": "BigScience", "country": "France", "date": "2022-11-03", "precision": "day", "category": "language", "open_weights": true, "params": "176B", "note": "Multitask-finetuned BLOOM, released with mT0, for crosslingual zero-shot", "source": "https://arxiv.org/abs/2211.01786" }, { "name": "Pangu-Weather", "org": "Huawei", "country": "China", "date": "2022-11-03", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "3D Earth-specific transformer for fast global weather forecasting", "source": "https://arxiv.org/abs/2211.02556" }, { "name": "Midjourney V4", "org": "Midjourney", "country": "USA", "date": "2022-11-05", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "New architecture and codebase; big leap in coherence", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "InternImage", "org": "Shanghai AI Lab", "country": "China", "date": "2022-11-10", "precision": "day", "category": "vision", "open_weights": true, "params": "1.08B", "note": "Deformable-convolution vision foundation model; COCO detection record", "source": "https://arxiv.org/abs/2211.05778" }, { "name": "AltCLIP", "org": "BAAI", "country": "China", "date": "2022-11-12", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "CLIP with a swapped-in multilingual text encoder", "source": "https://arxiv.org/abs/2211.06679" }, { "name": "EVA", "org": "BAAI", "country": "China", "date": "2022-11-14", "precision": "day", "category": "vision", "open_weights": true, "params": "1B", "note": "Billion-scale masked-image-modeling vision foundation model", "source": "https://arxiv.org/abs/2211.07636" }, { "name": "Galactica", "org": "Meta", "country": "USA", "date": "2022-11-15", "precision": "day", "category": "language", "open_weights": true, "params": "120B", "note": "Science LLM whose public demo was pulled after three days", "source": "https://arxiv.org/abs/2211.09085" }, { "name": "InstructPix2Pix", "org": "UC Berkeley", "country": "USA", "date": "2022-11-17", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Edits images by following written instructions", "source": "https://arxiv.org/abs/2211.09800" }, { "name": "Magic3D", "org": "NVIDIA", "country": "USA", "date": "2022-11-18", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "High-resolution text-to-3D, faster than DreamFusion", "source": "https://arxiv.org/abs/2211.10440" }, { "name": "MagicVideo", "org": "ByteDance", "country": "China", "date": "2022-11-20", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "ByteDance latent-diffusion text-to-video model", "source": "https://arxiv.org/abs/2211.11018" }, { "name": "CICERO", "org": "Meta", "country": "USA", "date": "2022-11-22", "precision": "day", "category": "games", "open_weights": true, "params": "", "note": "Human-level Diplomacy combining dialogue and strategic reasoning", "source": "https://ai.meta.com/blog/cicero-ai-negotiates-persuades-and-cooperates-with-people/" }, { "name": "Kandinsky 2.0", "org": "Sber", "country": "Russia", "date": "2022-11-23", "precision": "day", "category": "image", "open_weights": true, "params": "1.2B", "note": "Multilingual latent diffusion text-to-image model supporting 100+ languages", "source": "https://habr.com/ru/company/sberbank/blog/701162/" }, { "name": "Stable Diffusion 2.0", "org": "Stability AI", "country": "UK", "date": "2022-11-24", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Retrained from scratch with OpenCLIP encoder; 768x768 output", "source": "https://stability.ai/news-updates/stable-diffusion-v2-release" }, { "name": "Stable Diffusion 2.0 depth2img", "org": "Stability AI", "country": "UK", "date": "2022-11-24", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Depth-conditioned structure-preserving image-to-image model", "source": "https://stability.ai/news-updates/stable-diffusion-v2-release" }, { "name": "Stable Diffusion x4 Upscaler", "org": "Stability AI", "country": "UK", "date": "2022-11-24", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Text-guided 4x super-resolution diffusion model", "source": "https://stability.ai/news-updates/stable-diffusion-v2-release" }, { "name": "text-davinci-003", "org": "OpenAI", "country": "USA", "date": "2022-11-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Improved GPT-3.5 instruct model, two days before ChatGPT", "source": "https://en.wikipedia.org/wiki/GPT-3" }, { "name": "GPT-JT", "org": "Together AI", "country": "USA", "date": "2022-11-29", "precision": "day", "category": "language", "open_weights": true, "params": "6B", "note": "GPT-J fork fine-tuned with decentralized training over slow networks", "source": "https://www.together.ai/blog/releasing-v1-of-gpt-jt-powered-by-open-source-ai" }, { "name": "ChatGPT", "org": "OpenAI", "country": "USA", "date": "2022-11-30", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "GPT-3.5 chat research preview that ignited the generative AI boom", "source": "https://en.wikipedia.org/wiki/2022_in_artificial_intelligence" }, { "name": "Karlo", "org": "Kakao Brain", "country": "South Korea", "date": "2022-12-01", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open unCLIP-style text-to-image model (v1.0 alpha)", "source": "https://github.com/kakaobrain/karlo" }, { "name": "RFdiffusion", "org": "University of Washington", "country": "USA", "date": "2022-12-01", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Diffusion on RoseTTAFold for de novo protein design", "source": "https://www.ipd.uw.edu/2022/12/a-diffusion-model-for-protein-design/" }, { "name": "Chroma", "org": "Generate Biomedicines", "country": "USA", "date": "2022-12-02", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Programmable diffusion generative model for proteins", "source": "https://www.biorxiv.org/content/10.1101/2022.12.01.518682v1" }, { "name": "Whisper large-v2", "org": "OpenAI", "country": "USA", "date": "2022-12-05", "precision": "day", "category": "audio", "open_weights": true, "params": "1.55B", "note": "Retrained Whisper large with 2.5x more epochs and better accuracy", "source": "https://github.com/openai/whisper/commits/main/whisper/__init__.py" }, { "name": "E5", "org": "Microsoft", "country": "USA", "date": "2022-12-07", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Contrastive text embeddings; first to beat BM25 zero-shot on BEIR", "source": "https://arxiv.org/abs/2212.03533" }, { "name": "Stable Diffusion 2.1", "org": "Stability AI", "country": "UK", "date": "2022-12-07", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Fine-tuned from 2.0 with relaxed NSFW filtering", "source": "https://stability.ai/news/stablediffusion2-1-release7-dec-2022" }, { "name": "multilingual-22-12", "org": "Cohere", "country": "Canada", "date": "2022-12-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Cohere's first multilingual embedding model, covering 100+ languages", "source": "https://cohere.com/blog/multilingual" }, { "name": "data2vec 2.0", "org": "Meta", "country": "USA", "date": "2022-12-13", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Much faster self-supervised learning for vision, speech and text", "source": "https://ai.meta.com/blog/ai-self-supervised-learning-data2vec/" }, { "name": "Imagen Editor", "org": "Google", "country": "USA", "date": "2022-12-13", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Text-guided inpainting model, released with EditBench", "source": "https://arxiv.org/abs/2212.06909" }, { "name": "RT-1", "org": "Google", "country": "USA", "date": "2022-12-13", "precision": "day", "category": "robotics", "open_weights": true, "params": "35M", "note": "Robotics Transformer trained on 130k real-world robot episodes", "source": "https://arxiv.org/abs/2212.06817" }, { "name": "BioMedLM", "org": "Stanford University", "country": "USA", "date": "2022-12-15", "precision": "day", "category": "language", "open_weights": true, "params": "2.7B", "note": "Biomedical GPT (announced as PubMedGPT) trained with MosaicML", "source": "https://crfm.stanford.edu/2022/12/15/biomedlm.html" }, { "name": "Constitutional AI", "org": "Anthropic", "country": "USA", "date": "2022-12-15", "precision": "day", "category": "language", "open_weights": false, "params": "52B", "note": "Harmlessness trained from AI feedback against a written constitution", "source": "https://arxiv.org/abs/2212.08073" }, { "name": "Riffusion", "org": "Riffusion", "country": "USA", "date": "2022-12-15", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Stable Diffusion fine-tuned on spectrograms to generate music", "source": "https://en.wikipedia.org/wiki/Riffusion" }, { "name": "text-embedding-ada-002", "org": "OpenAI", "country": "USA", "date": "2022-12-15", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Single cheaper embedding model replacing five earlier OpenAI models", "source": "https://openai.com/index/new-and-improved-embedding-model/" }, { "name": "Point-E", "org": "OpenAI", "country": "USA", "date": "2022-12-16", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Fast text-to-3D point clouds on a single GPU", "source": "https://arxiv.org/abs/2212.08751" }, { "name": "DiT (Diffusion Transformer)", "org": "UC Berkeley", "country": "USA", "date": "2022-12-19", "precision": "day", "category": "image", "open_weights": true, "params": "675M", "note": "Transformer backbone for latent diffusion; template for later image/video models", "source": "https://arxiv.org/abs/2212.09748" }, { "name": "INSTRUCTOR", "org": "University of Hong Kong", "country": "China", "date": "2022-12-19", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Instruction-finetuned text embeddings adaptable to any task", "source": "https://arxiv.org/abs/2212.09741" }, { "name": "Niji V4 (niji・journey)", "org": "Midjourney", "country": "USA", "date": "2022-12-20", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "First anime and illustration-tuned Midjourney model, with Spellbrush", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "OPT-IML", "org": "Meta", "country": "USA", "date": "2022-12-22", "precision": "day", "category": "language", "open_weights": true, "params": "175B", "note": "OPT instruction-tuned on ~2,000 NLP tasks", "source": "https://arxiv.org/abs/2212.12017" }, { "name": "SantaCoder", "org": "BigCode", "country": "USA", "date": "2022-12-22", "precision": "day", "category": "code", "open_weights": true, "params": "1.1B", "note": "BigCode project's first open code model", "source": "https://huggingface.co/bigcode/santacoder" }, { "name": "Tune-A-Video", "org": "National University of Singapore", "country": "Singapore", "date": "2022-12-22", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "One-shot tuning of image diffusion for text-to-video, with Tencent ARC", "source": "https://arxiv.org/abs/2212.11565" }, { "name": "GraphCast", "org": "DeepMind", "country": "UK", "date": "2022-12-24", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Graph neural network for skillful medium-range weather forecasting", "source": "https://arxiv.org/abs/2212.12794" }, { "name": "Med-PaLM", "org": "Google", "country": "USA", "date": "2022-12-26", "precision": "day", "category": "language", "open_weights": false, "params": "540B", "note": "Flan-PaLM instruction-prompt-tuned for medical question answering", "source": "https://arxiv.org/abs/2212.13138" }, { "name": "H3 (Hungry Hungry Hippos)", "org": "Stanford University", "country": "USA", "date": "2022-12-28", "precision": "day", "category": "language", "open_weights": true, "params": "2.7B", "note": "State-space language model rivaling Transformers; precursor to Mamba", "source": "https://arxiv.org/abs/2212.14052" }, { "name": "Eleven Monolingual v1", "org": "ElevenLabs", "country": "USA", "date": "2023-01-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "ElevenLabs' first public TTS model, launched with its beta platform", "source": "https://en.wikipedia.org/wiki/ElevenLabs" }, { "name": "Muse", "org": "Google", "country": "USA", "date": "2023-01-02", "precision": "day", "category": "image", "open_weights": false, "params": "3B", "note": "Masked-token text-to-image transformer, faster than diffusion models", "source": "https://arxiv.org/abs/2301.00704" }, { "name": "VALL-E", "org": "Microsoft", "country": "USA", "date": "2023-01-05", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Neural codec language model cloning voices from 3-second samples", "source": "https://arxiv.org/abs/2301.02111" }, { "name": "DreamerV3", "org": "DeepMind", "country": "UK", "date": "2023-01-10", "precision": "day", "category": "games", "open_weights": false, "params": "200M", "note": "World-model agent; first to collect Minecraft diamonds from scratch", "source": "https://arxiv.org/abs/2301.04104" }, { "name": "YOLOv8", "org": "Ultralytics", "country": "USA", "date": "2023-01-10", "precision": "day", "category": "vision", "open_weights": true, "params": "3M-68M", "note": "Widely used real-time detection and segmentation model family", "source": "https://docs.ultralytics.com/models/yolov8/" }, { "name": "Nucleotide Transformer", "org": "InstaDeep", "country": "UK", "date": "2023-01-15", "precision": "day", "category": "science", "open_weights": true, "params": "2.5B", "note": "Open DNA foundation models built with NVIDIA and TU Munich", "source": "https://www.biorxiv.org/content/10.1101/2023.01.11.523679v1.full.pdf" }, { "name": "Adaptive Agent (AdA)", "org": "DeepMind", "country": "UK", "date": "2023-01-18", "precision": "day", "category": "games", "open_weights": false, "params": "533M", "note": "RL agent adapting to novel 3D tasks at human timescales", "source": "https://arxiv.org/abs/2301.07608" }, { "name": "I-JEPA", "org": "Meta", "country": "USA", "date": "2023-01-19", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "First model on LeCun's joint-embedding predictive architecture", "source": "https://arxiv.org/abs/2301.08243" }, { "name": "ClimaX", "org": "Microsoft", "country": "USA", "date": "2023-01-24", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Foundation model for weather and climate tasks", "source": "https://arxiv.org/abs/2301.10343" }, { "name": "MusicLM", "org": "Google", "country": "USA", "date": "2023-01-26", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "High-fidelity text-to-music generation at 24 kHz", "source": "https://arxiv.org/abs/2301.11325" }, { "name": "AudioLDM", "org": "University of Surrey", "country": "UK", "date": "2023-01-29", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open latent-diffusion text-to-audio generator", "source": "https://arxiv.org/abs/2301.12503" }, { "name": "BLIP-2", "org": "Salesforce", "country": "USA", "date": "2023-01-30", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Q-Former bridged frozen image encoders to frozen LLMs", "source": "https://arxiv.org/abs/2301.12597" }, { "name": "Bard", "org": "Google", "country": "USA", "date": "2023-02-06", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Google's rushed ChatGPT rival, initially powered by LaMDA", "source": "https://blog.google/technology/ai/bard-google-ai-search-updates/" }, { "name": "Gen-1", "org": "Runway", "country": "USA", "date": "2023-02-06", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Video-to-video generation guided by text or image prompts", "source": "https://arxiv.org/abs/2302.03011" }, { "name": "Bing Chat", "org": "Microsoft", "country": "USA", "date": "2023-02-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "New Bing chatbot; first public deployment of then-unannounced GPT-4", "source": "https://blogs.microsoft.com/blog/2023/02/07/reinventing-search-with-a-new-ai-powered-microsoft-bing-and-edge-your-copilot-for-the-web/" }, { "name": "Toolformer", "org": "Meta", "country": "USA", "date": "2023-02-09", "precision": "day", "category": "language", "open_weights": false, "params": "6.7B", "note": "Language model that teaches itself to call external tools", "source": "https://arxiv.org/abs/2302.04761" }, { "name": "ControlNet", "org": "Stanford University", "country": "USA", "date": "2023-02-10", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Added pose, edge and depth conditioning to Stable Diffusion", "source": "https://arxiv.org/abs/2302.05543" }, { "name": "ViT-22B", "org": "Google", "country": "USA", "date": "2023-02-10", "precision": "day", "category": "vision", "open_weights": false, "params": "22B", "note": "Largest dense vision transformer at the time", "source": "https://arxiv.org/abs/2302.05442" }, { "name": "MOSS", "org": "Fudan University", "country": "China", "date": "2023-02-20", "precision": "day", "category": "language", "open_weights": true, "params": "16B", "note": "China's first public ChatGPT-style chatbot; open-sourced in April", "source": "https://github.com/OpenMOSS/MOSS" }, { "name": "LLaMA", "org": "Meta", "country": "USA", "date": "2023-02-24", "precision": "day", "category": "language", "open_weights": true, "params": "7B-65B", "note": "Research weights leaked, igniting the open-weight LLM boom", "source": "https://ai.meta.com/blog/large-language-model-llama-meta-ai/" }, { "name": "Kosmos-1", "org": "Microsoft", "country": "USA", "date": "2023-02-27", "precision": "day", "category": "multimodal", "open_weights": false, "params": "1.6B", "note": "Early multimodal LLM perceiving interleaved images and text", "source": "https://arxiv.org/abs/2302.14045" }, { "name": "GPT-3.5 Turbo", "org": "OpenAI", "country": "USA", "date": "2023-03-01", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "ChatGPT's model via API at one-tenth prior GPT-3.5 price", "source": "https://openai.com/index/introducing-chatgpt-and-whisper-apis/" }, { "name": "Consistency Models", "org": "OpenAI", "country": "USA", "date": "2023-03-02", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "One-step generation technique distilled from diffusion models", "source": "https://arxiv.org/abs/2303.01469" }, { "name": "Universal Speech Model (USM)", "org": "Google", "country": "USA", "date": "2023-03-02", "precision": "day", "category": "audio", "open_weights": false, "params": "2B", "note": "Speech recognition model spanning 300+ languages", "source": "https://arxiv.org/abs/2303.01037" }, { "name": "Flan-UL2", "org": "Google", "country": "USA", "date": "2023-03-03", "precision": "day", "category": "language", "open_weights": true, "params": "20B", "note": "Apache-licensed 20B instruction-tuned UL2 model", "source": "https://www.yitay.net/blog/flan-ul2-20b" }, { "name": "PaLM-E", "org": "Google", "country": "USA", "date": "2023-03-06", "precision": "day", "category": "robotics", "open_weights": false, "params": "562B", "note": "Embodied multimodal LLM planning robot manipulation", "source": "https://arxiv.org/abs/2303.03378" }, { "name": "Diffusion Policy", "org": "Columbia University", "country": "USA", "date": "2023-03-07", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Diffusion models applied to visuomotor robot policy learning", "source": "https://arxiv.org/abs/2303.04137" }, { "name": "VALL-E X", "org": "Microsoft", "country": "USA", "date": "2023-03-07", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Cross-lingual zero-shot voice cloning speech synthesis", "source": "https://arxiv.org/abs/2303.03926" }, { "name": "Grounding DINO", "org": "IDEA Research", "country": "China", "date": "2023-03-09", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Open-set object detection from free-text prompts", "source": "https://arxiv.org/abs/2303.05499" }, { "name": "Jurassic-2", "org": "AI21 Labs", "country": "Israel", "date": "2023-03-09", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "AI21's second-generation LLM family with task-specific APIs", "source": "https://www.ai21.com/blog/introducing-j2/" }, { "name": "Alpaca", "org": "Stanford University", "country": "USA", "date": "2023-03-13", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "LLaMA instruction-tuned for under $600, sparking fine-tune wave", "source": "https://crfm.stanford.edu/2023/03/13/alpaca.html" }, { "name": "ChatGLM-6B", "org": "Zhipu AI", "country": "China", "date": "2023-03-14", "precision": "day", "category": "language", "open_weights": true, "params": "6B", "note": "Hugely popular open bilingual Chinese-English chat model", "source": "https://github.com/zai-org/ChatGLM-6B" }, { "name": "Claude", "org": "Anthropic", "country": "USA", "date": "2023-03-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Anthropic's first public model; context expanded to 100K in May", "source": "https://www.anthropic.com/news/introducing-claude" }, { "name": "Claude Instant", "org": "Anthropic", "country": "USA", "date": "2023-03-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Lighter, cheaper, faster Claude variant launched alongside Claude", "source": "https://www.anthropic.com/news/introducing-claude" }, { "name": "GPT-4", "org": "OpenAI", "country": "USA", "date": "2023-03-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Image-input frontier model with human-level professional exam scores", "source": "https://openai.com/index/gpt-4/" }, { "name": "Med-PaLM 2", "org": "Google", "country": "USA", "date": "2023-03-14", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "First model at expert level on USMLE-style medical questions", "source": "https://sites.research.google/gr/med-palm/" }, { "name": "Falcon-40B", "org": "TII", "country": "UAE", "date": "2023-03-15", "precision": "day", "category": "language", "open_weights": true, "params": "40B", "note": "Announced March; weights (with Falcon-7B) opened May, topped open leaderboard", "source": "https://www.tii.ae/news/abu-dhabi-based-technology-innovation-institute-introduces-falcon-llm-foundational-large" }, { "name": "Midjourney V5", "org": "Midjourney", "country": "USA", "date": "2023-03-15", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Photorealism leap behind viral Pope puffer-jacket fake", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "ERNIE Bot", "org": "Baidu", "country": "China", "date": "2023-03-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "First big-tech Chinese ChatGPT rival; invite-only launch", "source": "https://en.wikipedia.org/wiki/Ernie_Bot" }, { "name": "ModelScope Text-to-Video", "org": "Alibaba", "country": "China", "date": "2023-03-19", "precision": "day", "category": "video", "open_weights": true, "params": "1.7B", "note": "Early open-source text-to-video diffusion model from DAMO Academy", "source": "https://huggingface.co/ali-vilab/modelscope-damo-text-to-video-synthesis" }, { "name": "Gen-2", "org": "Runway", "country": "USA", "date": "2023-03-20", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Among the first commercially available text-to-video models", "source": "https://runway.com/research/gen-2" }, { "name": "PanGu-Σ", "org": "Huawei", "country": "China", "date": "2023-03-20", "precision": "day", "category": "language", "open_weights": false, "params": "1.085T", "note": "Trillion-parameter sparse LLM trained on Ascend chips", "source": "https://arxiv.org/abs/2303.10845" }, { "name": "Zero-1-to-3", "org": "Columbia University", "country": "USA", "date": "2023-03-20", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Single image to novel views and 3D", "source": "https://arxiv.org/abs/2303.11328" }, { "name": "Firefly", "org": "Adobe", "country": "USA", "date": "2023-03-21", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Image generator trained on licensed and public-domain content", "source": "https://news.adobe.com/news/news-details/2023/Adobe-Unveils-Firefly-a-Family-of-new-Creative-Generative-AI/default.aspx" }, { "name": "Dolly", "org": "Databricks", "country": "USA", "date": "2023-03-24", "precision": "day", "category": "language", "open_weights": true, "params": "6B", "note": "Showed cheap instruction tuning revives an old open model", "source": "https://www.databricks.com/blog/2023/03/24/hello-dolly-democratizing-magic-chatgpt-open-models.html" }, { "name": "SigLIP", "org": "Google", "country": "USA", "date": "2023-03-27", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Sigmoid-loss image-text pretraining; widely reused vision encoder", "source": "https://arxiv.org/abs/2303.15343" }, { "name": "Cerebras-GPT", "org": "Cerebras", "country": "USA", "date": "2023-03-28", "precision": "day", "category": "language", "open_weights": true, "params": "111M-13B", "note": "Compute-optimal open GPT family trained on wafer-scale hardware", "source": "https://arxiv.org/abs/2304.03208" }, { "name": "OpenFlamingo", "org": "LAION", "country": "Germany", "date": "2023-03-28", "precision": "day", "category": "multimodal", "open_weights": true, "params": "9B", "note": "Open reproduction of DeepMind's Flamingo vision-language model", "source": "https://laion.ai/blog/open-flamingo/" }, { "name": "BloombergGPT", "org": "Bloomberg", "country": "USA", "date": "2023-03-30", "precision": "day", "category": "language", "open_weights": false, "params": "50B", "note": "Domain-specific LLM trained on financial data", "source": "https://arxiv.org/abs/2303.17564" }, { "name": "Vicuna", "org": "LMSYS", "country": "USA", "date": "2023-03-30", "precision": "day", "category": "language", "open_weights": true, "params": "7B-13B", "note": "ShareGPT-tuned LLaMA claiming 90% of ChatGPT quality", "source": "https://lmsys.org/blog/2023-03-30-vicuna/" }, { "name": "Niji V5", "org": "Midjourney", "country": "USA", "date": "2023-04-02", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Anime-tuned model built on Midjourney V5", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "Koala", "org": "UC Berkeley", "country": "USA", "date": "2023-04-03", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Academic dialogue model fine-tuned from LLaMA on web data", "source": "https://bair.berkeley.edu/blog/2023/04/03/koala/" }, { "name": "Pythia", "org": "EleutherAI", "country": "USA", "date": "2023-04-03", "precision": "day", "category": "language", "open_weights": true, "params": "70M-12B", "note": "Fully open model suite for studying training dynamics", "source": "https://arxiv.org/abs/2304.01373" }, { "name": "Kandinsky 2.1", "org": "Sber", "country": "Russia", "date": "2023-04-04", "precision": "day", "category": "image", "open_weights": true, "params": "3.3B", "note": "Widely used Russian open text-to-image model", "source": "https://github.com/ai-forever/Kandinsky-2" }, { "name": "Segment Anything Model (SAM)", "org": "Meta", "country": "USA", "date": "2023-04-05", "precision": "day", "category": "vision", "open_weights": true, "params": "636M", "note": "Promptable segmentation foundation model with 1B-mask dataset", "source": "https://ai.meta.com/blog/segment-anything-foundation-model-image-segmentation/" }, { "name": "FengWu", "org": "Shanghai AI Lab", "country": "China", "date": "2023-04-06", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "AI global forecast pushing skillful lead time beyond 10 days", "source": "https://arxiv.org/abs/2304.02948" }, { "name": "SegGPT", "org": "BAAI", "country": "China", "date": "2023-04-06", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Segment-everything-in-context generalist segmentation model", "source": "https://arxiv.org/abs/2304.03284" }, { "name": "Generative Agents", "org": "Stanford University", "country": "USA", "date": "2023-04-07", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "LLM-driven simulated townsfolk with memory, reflection and planning", "source": "https://arxiv.org/abs/2304.03442" }, { "name": "Tongyi Qianwen", "org": "Alibaba", "country": "China", "date": "2023-04-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Alibaba's first ChatGPT-style LLM, later branded Qwen", "source": "https://zh.wikipedia.org/wiki/通义千问" }, { "name": "SenseNova (SenseChat)", "org": "SenseTime", "country": "China", "date": "2023-04-10", "precision": "day", "category": "language", "open_weights": false, "params": "180B", "note": "SenseTime foundation model suite and ChatGPT-style chatbot", "source": "https://www.reuters.com/technology/chinese-ai-firm-sensetime-unveils-chatbot-sensechat-2023-04-10/" }, { "name": "Dolly 2.0", "org": "Databricks", "country": "USA", "date": "2023-04-12", "precision": "day", "category": "language", "open_weights": true, "params": "12B", "note": "First open instruction-tuned LLM licensed for commercial use", "source": "https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm" }, { "name": "Amazon Titan", "org": "Amazon", "country": "USA", "date": "2023-04-13", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Amazon's foundation models unveiled alongside Bedrock", "source": "https://aws.amazon.com/blogs/machine-learning/announcing-new-tools-for-building-with-generative-ai-on-aws/" }, { "name": "SEEM", "org": "University of Wisconsin-Madison", "country": "USA", "date": "2023-04-13", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Segment everything everywhere with multimodal prompts", "source": "https://arxiv.org/abs/2304.06718" }, { "name": "DINOv2", "org": "Meta", "country": "USA", "date": "2023-04-14", "precision": "day", "category": "vision", "open_weights": true, "params": "1.1B", "note": "Self-supervised universal visual features", "source": "https://arxiv.org/abs/2304.07193" }, { "name": "OpenAssistant", "org": "LAION", "country": "Germany", "date": "2023-04-14", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Crowdsourced open chat assistant and conversation dataset", "source": "https://arxiv.org/abs/2304.07327" }, { "name": "LLaVA", "org": "University of Wisconsin-Madison", "country": "USA", "date": "2023-04-17", "precision": "day", "category": "multimodal", "open_weights": true, "params": "13B", "note": "Pioneered visual instruction tuning for open vision-language models", "source": "https://arxiv.org/abs/2304.08485" }, { "name": "NaturalSpeech 2", "org": "Microsoft", "country": "USA", "date": "2023-04-18", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Latent diffusion zero-shot speech and singing synthesis", "source": "https://arxiv.org/abs/2304.09116" }, { "name": "Video LDM", "org": "NVIDIA", "country": "USA", "date": "2023-04-18", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "High-resolution video synthesis with latent diffusion", "source": "https://arxiv.org/abs/2304.08818" }, { "name": "StableLM", "org": "Stability AI", "country": "UK", "date": "2023-04-19", "precision": "day", "category": "language", "open_weights": true, "params": "3B-7B", "note": "Stability AI's first open language model suite", "source": "https://stability.ai/news-updates/stability-ai-launches-the-first-of-its-stablelm-suite-of-language-models" }, { "name": "Bark", "org": "Suno", "country": "USA", "date": "2023-04-20", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open text-to-audio model for speech, music and effects", "source": "https://github.com/suno-ai/bark" }, { "name": "MiniGPT-4", "org": "KAUST", "country": "Saudi Arabia", "date": "2023-04-20", "precision": "day", "category": "multimodal", "open_weights": true, "params": "13B", "note": "Open reproduction of GPT-4-style image understanding", "source": "https://arxiv.org/abs/2304.10592" }, { "name": "ACT (ALOHA)", "org": "Stanford University", "country": "USA", "date": "2023-04-23", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Low-cost bimanual robot learning fine manipulation from demos", "source": "https://arxiv.org/abs/2304.13705" }, { "name": "GigaChat", "org": "Sber", "country": "Russia", "date": "2023-04-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Sberbank's Russian alternative to ChatGPT, launched in closed testing", "source": "https://en.wikipedia.org/wiki/GigaChat" }, { "name": "WizardLM", "org": "Microsoft", "country": "USA", "date": "2023-04-24", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Evol-Instruct synthetic data for complex instruction following", "source": "https://arxiv.org/abs/2304.12244" }, { "name": "Eleven Multilingual v1", "org": "ElevenLabs", "country": "USA", "date": "2023-04-27", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "ElevenLabs' first multilingual speech synthesis model", "source": "https://elevenlabs.io/blog/eleven-multilingual-v1" }, { "name": "mPLUG-Owl", "org": "Alibaba", "country": "China", "date": "2023-04-27", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "DAMO Academy's modular open multimodal LLM", "source": "https://arxiv.org/abs/2304.14178" }, { "name": "DeepFloyd IF", "org": "Stability AI", "country": "UK", "date": "2023-04-28", "precision": "day", "category": "image", "open_weights": true, "params": "4.3B", "note": "Pixel-space diffusion with strong in-image text rendering", "source": "https://stability.ai/news-updates/deepfloyd-if-text-to-image-model" }, { "name": "OpenLLaMA", "org": "OpenLM Research", "country": "USA", "date": "2023-05-01", "precision": "month", "category": "language", "open_weights": true, "params": "3B-13B", "note": "Permissively licensed open reproduction of LLaMA", "source": "https://github.com/openlm-research/open_llama" }, { "name": "Midjourney V5.1", "org": "Midjourney", "country": "USA", "date": "2023-05-03", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "More opinionated default aesthetic and improved coherence", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "Shap-E", "org": "OpenAI", "country": "USA", "date": "2023-05-03", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open text-to-3D and image-to-3D generator", "source": "https://arxiv.org/abs/2305.02463" }, { "name": "StarCoder", "org": "Hugging Face", "country": "USA", "date": "2023-05-04", "precision": "day", "category": "code", "open_weights": true, "params": "15.5B", "note": "BigCode open code LLM built with ServiceNow on permissive code", "source": "https://huggingface.co/blog/starcoder" }, { "name": "MPT-7B", "org": "MosaicML", "country": "USA", "date": "2023-05-05", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Commercially usable open LLM trained on 1T tokens", "source": "https://www.databricks.com/blog/mpt-7b" }, { "name": "RedPajama-INCITE", "org": "Together AI", "country": "USA", "date": "2023-05-05", "precision": "day", "category": "language", "open_weights": true, "params": "3B-7B", "note": "Open models trained on a reproduction of LLaMA's dataset", "source": "https://www.together.ai/blog/redpajama-models-v1" }, { "name": "SparkDesk", "org": "iFlytek", "country": "China", "date": "2023-05-06", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "iFlytek's Spark LLM debut, claimed to rival ChatGPT", "source": "https://technode.com/2023/05/08/iflytek-unveils-large-language-model-claims-it-outperforms-chatgpt/" }, { "name": "ImageBind", "org": "Meta", "country": "USA", "date": "2023-05-09", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Single embedding space binding six modalities", "source": "https://arxiv.org/abs/2305.05665" }, { "name": "Codey", "org": "Google", "country": "USA", "date": "2023-05-10", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "PaLM 2-based code generation model on Vertex AI", "source": "https://blog.google/technology/developers/google-io-2023-100-announcements" }, { "name": "PaLM 2", "org": "Google", "country": "USA", "date": "2023-05-10", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Powered Bard and 25 products; sizes Gecko to Unicorn", "source": "https://blog.google/technology/ai/google-palm-2-ai-large-language-model/" }, { "name": "InstructBLIP", "org": "Salesforce", "country": "USA", "date": "2023-05-11", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Instruction-tuned general-purpose vision-language models", "source": "https://arxiv.org/abs/2305.06500" }, { "name": "CodeT5+", "org": "Salesforce", "country": "USA", "date": "2023-05-13", "precision": "day", "category": "code", "open_weights": true, "params": "220M-16B", "note": "Open encoder-decoder code LLM family", "source": "https://arxiv.org/abs/2305.07922" }, { "name": "SoundStorm", "org": "Google", "country": "USA", "date": "2023-05-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Parallel audio generation 100x faster than AudioLM", "source": "https://arxiv.org/abs/2305.09636" }, { "name": "VisualGLM-6B", "org": "Zhipu AI", "country": "China", "date": "2023-05-17", "precision": "day", "category": "multimodal", "open_weights": true, "params": "7.8B", "note": "Open ChatGLM-based image-understanding chat model", "source": "https://github.com/zai-org/ChatGLM-6B" }, { "name": "YandexGPT", "org": "Yandex", "country": "Russia", "date": "2023-05-17", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Yandex's LLM added to the Alice voice assistant", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "DragGAN", "org": "Max Planck Institute for Informatics", "country": "Germany", "date": "2023-05-18", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Viral point-drag image editing on the GAN manifold", "source": "https://arxiv.org/abs/2305.10973" }, { "name": "MMS (Massively Multilingual Speech)", "org": "Meta", "country": "USA", "date": "2023-05-22", "precision": "day", "category": "audio", "open_weights": true, "params": "1B", "note": "Speech recognition and synthesis for 1,100+ languages", "source": "https://arxiv.org/abs/2305.13516" }, { "name": "Guanaco (QLoRA)", "org": "University of Washington", "country": "USA", "date": "2023-05-23", "precision": "day", "category": "language", "open_weights": true, "params": "7B-65B", "note": "4-bit finetuning trained a 65B chat model on one GPU", "source": "https://arxiv.org/abs/2305.14314" }, { "name": "Gorilla", "org": "UC Berkeley", "country": "USA", "date": "2023-05-24", "precision": "day", "category": "agent", "open_weights": true, "params": "7B", "note": "LLM fine-tuned to write accurate API calls", "source": "https://arxiv.org/abs/2305.15334" }, { "name": "Voyager", "org": "NVIDIA", "country": "USA", "date": "2023-05-25", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "GPT-4-powered lifelong-learning agent exploring Minecraft", "source": "https://arxiv.org/abs/2305.16291" }, { "name": "PaLI-X", "org": "Google", "country": "USA", "date": "2023-05-29", "precision": "day", "category": "multimodal", "open_weights": false, "params": "55B", "note": "Scaled multilingual vision-language model for images, documents and video", "source": "https://arxiv.org/abs/2305.18565" }, { "name": "Orca", "org": "Microsoft", "country": "USA", "date": "2023-06-05", "precision": "day", "category": "language", "open_weights": false, "params": "13B", "note": "Small model imitating GPT-4 explanation traces", "source": "https://arxiv.org/abs/2306.02707" }, { "name": "LTM-1", "org": "Magic", "country": "USA", "date": "2023-06-06", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Code model with a 5-million-token context window", "source": "https://magic.dev/blog/ltm-1" }, { "name": "MetNet-3", "org": "Google", "country": "USA", "date": "2023-06-06", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Neural weather model beating physics models to 24 hours", "source": "https://arxiv.org/abs/2306.06079" }, { "name": "AlphaDev", "org": "Google DeepMind", "country": "UK", "date": "2023-06-07", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "RL-discovered sorting algorithms merged into C++ standard library", "source": "https://deepmind.google/blog/alphadev-discovers-faster-sorting-algorithms/" }, { "name": "InternLM", "org": "Shanghai AI Lab", "country": "China", "date": "2023-06-07", "precision": "day", "category": "language", "open_weights": false, "params": "104B", "note": "104B multilingual model built with SenseTime, trained on 1.6T tokens", "source": "https://github.com/InternLM/InternLM-techreport" }, { "name": "Tülu", "org": "Allen Institute for AI", "country": "USA", "date": "2023-06-07", "precision": "day", "category": "language", "open_weights": true, "params": "7B-65B", "note": "First Ai2 open instruction-tuning suite on open base models", "source": "https://arxiv.org/abs/2306.04751" }, { "name": "MusicGen", "org": "Meta", "country": "USA", "date": "2023-06-08", "precision": "day", "category": "audio", "open_weights": true, "params": "3.3B", "note": "Open single-stage controllable text-to-music model", "source": "https://arxiv.org/abs/2306.05284" }, { "name": "Aquila", "org": "BAAI", "country": "China", "date": "2023-06-09", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "BAAI's bilingual open LLM (Wu Dao 3.0) with commercial license", "source": "https://spectrum.ieee.org/china-chatgpt-wu-dao" }, { "name": "WizardCoder", "org": "Microsoft", "country": "USA", "date": "2023-06-14", "precision": "day", "category": "code", "open_weights": true, "params": "15B", "note": "Evol-Instruct code model beating Claude and Bard on HumanEval", "source": "https://arxiv.org/abs/2306.08568" }, { "name": "Baichuan-7B", "org": "Baichuan", "country": "China", "date": "2023-06-15", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Baichuan's first open model, two months after founding", "source": "https://huggingface.co/baichuan-inc/Baichuan-7B" }, { "name": "Voicebox", "org": "Meta", "country": "USA", "date": "2023-06-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Flow-matching speech generator withheld over misuse concerns", "source": "https://about.fb.com/news/2023/06/introducing-voicebox-ai-for-speech-generation/" }, { "name": "GAIA-1", "org": "Wayve", "country": "UK", "date": "2023-06-17", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Generative world model for autonomous driving video", "source": "https://wayve.ai/thinking/introducing-gaia1/" }, { "name": "phi-1", "org": "Microsoft", "country": "USA", "date": "2023-06-20", "precision": "day", "category": "code", "open_weights": true, "params": "1.3B", "note": "'Textbooks Are All You Need' small high-quality code model", "source": "https://arxiv.org/abs/2306.11644" }, { "name": "RoboCat", "org": "Google DeepMind", "country": "UK", "date": "2023-06-20", "precision": "day", "category": "robotics", "open_weights": false, "params": "1.18B", "note": "Self-improving agent generalizing across different robot arms", "source": "https://deepmind.google/blog/robocat-a-self-improving-robotic-agent/" }, { "name": "AudioPaLM", "org": "Google", "country": "USA", "date": "2023-06-22", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "PaLM 2 fused with AudioLM for speech understanding and translation", "source": "https://arxiv.org/abs/2306.12925" }, { "name": "FuXi", "org": "Fudan University", "country": "China", "date": "2023-06-22", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Cascaded ML system for 15-day global weather forecasts", "source": "https://arxiv.org/abs/2306.12873" }, { "name": "Inflection-1", "org": "Inflection AI", "country": "USA", "date": "2023-06-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "In-house model powering the Pi personal assistant", "source": "https://inflection.ai/assets/Inflection-1.pdf" }, { "name": "Midjourney V5.2", "org": "Midjourney", "country": "USA", "date": "2023-06-22", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Added zoom-out outpainting and new aesthetics system", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "MPT-30B", "org": "MosaicML", "country": "USA", "date": "2023-06-22", "precision": "day", "category": "language", "open_weights": true, "params": "30B", "note": "Open 30B model with 8K context, commercially licensed", "source": "https://www.databricks.com/blog/mpt-30b" }, { "name": "SDXL 0.9", "org": "Stability AI", "country": "UK", "date": "2023-06-22", "precision": "day", "category": "image", "open_weights": true, "params": "3.5B", "note": "Research preview of Stable Diffusion XL", "source": "https://stability.ai/news-updates/sdxl-09-stable-diffusion" }, { "name": "ChatGLM2-6B", "org": "Zhipu AI", "country": "China", "date": "2023-06-25", "precision": "day", "category": "language", "open_weights": true, "params": "6B", "note": "Upgraded open bilingual chat model with 32K context", "source": "https://github.com/zai-org/ChatGLM-6B" }, { "name": "Kosmos-2", "org": "Microsoft", "country": "USA", "date": "2023-06-26", "precision": "day", "category": "multimodal", "open_weights": true, "params": "1.6B", "note": "Grounded multimodal LLM linking text to image regions", "source": "https://arxiv.org/abs/2306.14824" }, { "name": "ERNIE 3.5", "org": "Baidu", "country": "China", "date": "2023-06-27", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Upgraded ERNIE claimed to beat ChatGPT on Chinese benchmarks", "source": "http://research.baidu.com/Blog/index-view?id=185" }, { "name": "HyenaDNA", "org": "Stanford University", "country": "USA", "date": "2023-06-27", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Genomic model with 1M-token context at single-nucleotide resolution", "source": "https://arxiv.org/abs/2306.15794" }, { "name": "XGen-7B", "org": "Salesforce", "country": "USA", "date": "2023-06-28", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Open 7B LLM trained with 8K context", "source": "https://huggingface.co/Salesforce/xgen-7b-8k-base" }, { "name": "InternLM-7B", "org": "Shanghai AI Lab", "country": "China", "date": "2023-07-06", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "First open-weight InternLM release, unveiled at WAIC", "source": "https://huggingface.co/internlm/internlm-7b" }, { "name": "xTrimoPGLM", "org": "BioMap", "country": "China", "date": "2023-07-06", "precision": "day", "category": "science", "open_weights": false, "params": "100B", "note": "100B-parameter protein language model with Tsinghua", "source": "https://www.biorxiv.org/content/10.1101/2023.07.05.547496v4" }, { "name": "Pangu 3.0", "org": "Huawei", "country": "China", "date": "2023-07-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Industry-focused foundation model family on Huawei Cloud", "source": "https://www.huaweicloud.com/intl/en-us/news/20230707180809498.html" }, { "name": "AnimateDiff", "org": "Shanghai AI Lab", "country": "China", "date": "2023-07-10", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Plug-in motion module animating personalized text-to-image models", "source": "https://arxiv.org/abs/2307.04725" }, { "name": "Baichuan-13B", "org": "Baichuan", "country": "China", "date": "2023-07-11", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Open 13B bilingual model with commercial license", "source": "https://github.com/baichuan-inc/Baichuan-13B" }, { "name": "Claude 2", "org": "Anthropic", "country": "USA", "date": "2023-07-11", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "100K context; launched public claude.ai beta", "source": "https://www.anthropic.com/news/claude-2" }, { "name": "Emu", "org": "BAAI", "country": "China", "date": "2023-07-11", "precision": "day", "category": "multimodal", "open_weights": true, "params": "14B", "note": "Generative multimodal model predicting interleaved images and text", "source": "https://arxiv.org/abs/2307.05222" }, { "name": "Kandinsky 2.2", "org": "Sber", "country": "Russia", "date": "2023-07-12", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Upgraded open Kandinsky with larger CLIP encoder and ControlNet support", "source": "https://github.com/ai-forever/Kandinsky-2" }, { "name": "ChatRhino", "org": "JD.com", "country": "China", "date": "2023-07-13", "precision": "day", "category": "language", "open_weights": false, "params": "100B", "note": "JD's industry LLM for retail, logistics and health", "source": "https://jdcorporateblog.com/jd-com-introduces-chatrhino-empowering-industry-innovations-with-an-advanced-large-language-model/" }, { "name": "CM3leon", "org": "Meta", "country": "USA", "date": "2023-07-14", "precision": "day", "category": "image", "open_weights": false, "params": "7B", "note": "Single autoregressive model for text-to-image and captioning", "source": "https://ai.meta.com/blog/generative-ai-text-images-cm3leon/" }, { "name": "RetNet", "org": "Microsoft", "country": "USA", "date": "2023-07-17", "precision": "day", "category": "language", "open_weights": false, "params": "6.7B", "note": "Retentive network proposed as a transformer successor", "source": "https://arxiv.org/abs/2307.08621" }, { "name": "Llama 2", "org": "Meta", "country": "USA", "date": "2023-07-18", "precision": "day", "category": "language", "open_weights": true, "params": "7B-70B", "note": "Commercially licensed open weights, launched with Microsoft", "source": "https://about.fb.com/news/2023/07/llama-2/" }, { "name": "EXAONE 2.0", "org": "LG AI Research", "country": "South Korea", "date": "2023-07-19", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "LG's expert-domain multimodal foundation model", "source": "https://www.lgresearch.ai/exaone" }, { "name": "Stable Beluga 2", "org": "Stability AI", "country": "UK", "date": "2023-07-21", "precision": "day", "category": "language", "open_weights": true, "params": "70B", "note": "Orca-style Llama 2 70B fine-tune, first announced as FreeWilly2", "source": "https://stability.ai/news-updates/stable-beluga-large-instruction-fine-tuned-models" }, { "name": "BTLM-3B-8K", "org": "Cerebras", "country": "USA", "date": "2023-07-24", "precision": "day", "category": "language", "open_weights": true, "params": "2.6B", "note": "3B model matching 7B quality, 8K context", "source": "https://arxiv.org/abs/2309.11568" }, { "name": "CodeGeeX2", "org": "Zhipu AI", "country": "China", "date": "2023-07-25", "precision": "day", "category": "code", "open_weights": true, "params": "6B", "note": "ChatGLM2-based multilingual code generation model", "source": "https://github.com/zai-org/ChatGLM-6B" }, { "name": "Med-PaLM M", "org": "Google", "country": "USA", "date": "2023-07-26", "precision": "day", "category": "science", "open_weights": false, "params": "562B", "note": "Generalist biomedical model across imaging, genomics and text", "source": "https://arxiv.org/abs/2307.14334" }, { "name": "SDXL 1.0", "org": "Stability AI", "country": "UK", "date": "2023-07-26", "precision": "day", "category": "image", "open_weights": true, "params": "3.5B", "note": "Flagship open 1024px text-to-image model", "source": "https://stability.ai/news-updates/stable-diffusion-sdxl-1-announcement" }, { "name": "RT-2", "org": "Google DeepMind", "country": "UK", "date": "2023-07-28", "precision": "day", "category": "robotics", "open_weights": false, "params": "55B", "note": "First vision-language-action model transferring web knowledge to robots", "source": "https://deepmind.google/blog/rt-2-new-model-translates-vision-and-language-into-action/" }, { "name": "AudioCraft", "org": "Meta", "country": "USA", "date": "2023-08-02", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open release of MusicGen, AudioGen and improved EnCodec", "source": "https://about.fb.com/news/2023/08/audiocraft-generative-ai-for-music-and-audio/" }, { "name": "BGE", "org": "BAAI", "country": "China", "date": "2023-08-02", "precision": "day", "category": "language", "open_weights": true, "params": "335M", "note": "Open text-embedding family that debuted first on MTEB", "source": "https://github.com/FlagOpen/FlagEmbedding" }, { "name": "Prithvi", "org": "IBM", "country": "USA", "date": "2023-08-03", "precision": "day", "category": "science", "open_weights": true, "params": "100M", "note": "Largest open geospatial foundation model, built with NASA", "source": "https://newsroom.ibm.com/2023-08-03-IBM-and-NASA-Open-Source-Largest-Geospatial-AI-Foundation-Model-on-Hugging-Face" }, { "name": "Qwen-7B", "org": "Alibaba", "country": "China", "date": "2023-08-03", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "First open-weight Qwen release", "source": "https://github.com/QwenLM/Qwen" }, { "name": "GTE", "org": "Alibaba", "country": "China", "date": "2023-08-07", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "General Text Embeddings; small model rivaling OpenAI embedding API", "source": "https://arxiv.org/abs/2308.03281" }, { "name": "XVERSE-13B", "org": "XVERSE", "country": "China", "date": "2023-08-07", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Shenzhen startup's open multilingual base model", "source": "https://github.com/xverse-ai/XVERSE-13B" }, { "name": "Baichuan-53B", "org": "Baichuan", "country": "China", "date": "2023-08-08", "precision": "day", "category": "language", "open_weights": false, "params": "53B", "note": "Closed model; Baichuan's third LLM in four months", "source": "https://technode.com/2023/08/09/chinese-ai-startup-baichuan-rolls-out-third-llm-in-four-months/" }, { "name": "StableCode", "org": "Stability AI", "country": "UK", "date": "2023-08-08", "precision": "day", "category": "code", "open_weights": true, "params": "3B", "note": "Stability AI's first code LLM", "source": "https://stability.ai/news-updates/stablecode-llm-generative-ai-coding" }, { "name": "Claude Instant 1.2", "org": "Anthropic", "country": "USA", "date": "2023-08-09", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Cheaper fast model inheriting Claude 2 improvements", "source": "https://www.anthropic.com/news/releasing-claude-instant-1-2" }, { "name": "Japanese StableLM Alpha", "org": "Stability AI", "country": "UK", "date": "2023-08-10", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Japanese-focused open LLM trained on 750B tokens", "source": "https://huggingface.co/stabilityai/japanese-stablelm-base-alpha-7b" }, { "name": "SparkDesk V2.0", "org": "iFlytek", "country": "China", "date": "2023-08-15", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Spark upgrade adding multimodal and coding (iFlyCode) abilities", "source": "https://technode.com/2023/08/16/iflytek-unveils-updated-llm-sparkdesk-v2-0-and-new-product-iflycode-1-0/" }, { "name": "KwaiYii", "org": "Kuaishou", "country": "China", "date": "2023-08-16", "precision": "day", "category": "language", "open_weights": false, "params": "13B", "note": "Kuaishou's first self-developed LLM, detailed on GitHub", "source": "https://github.com/kwai/KwaiYii" }, { "name": "Doubao", "org": "ByteDance", "country": "China", "date": "2023-08-18", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "ByteDance's first chatbot, built on its in-house Skylark LLM", "source": "https://voicebot.ai/2023/08/18/tiktok-parent-company-bytedance-releases-generative-ai-chatbot-dou-bao-in-china/" }, { "name": "Eleven Multilingual v2", "org": "ElevenLabs", "country": "USA", "date": "2023-08-22", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech synthesis across 29 languages; platform left beta", "source": "https://elevenlabs.io/blog/eleven-multilingual-v2" }, { "name": "IDEFICS", "org": "Hugging Face", "country": "USA", "date": "2023-08-22", "precision": "day", "category": "language", "open_weights": true, "params": "9B-80B", "note": "Open reproduction of DeepMind's Flamingo vision-language model", "source": "https://huggingface.co/blog/idefics" }, { "name": "Ideogram 0.1", "org": "Ideogram", "country": "Canada", "date": "2023-08-22", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Image generator notable for reliable in-image typography", "source": "https://en.wikipedia.org/wiki/Ideogram_(text-to-image_model)" }, { "name": "Qwen-VL", "org": "Alibaba", "country": "China", "date": "2023-08-22", "precision": "day", "category": "language", "open_weights": true, "params": "9.6B", "note": "Open vision-language model with grounding and text reading", "source": "https://github.com/QwenLM/Qwen-VL" }, { "name": "SeamlessM4T", "org": "Meta", "country": "USA", "date": "2023-08-22", "precision": "day", "category": "audio", "open_weights": true, "params": "2.3B", "note": "All-in-one multilingual speech and text translation", "source": "https://about.fb.com/news/2023/08/seamlessm4t-ai-translation-model/" }, { "name": "Code Llama", "org": "Meta", "country": "USA", "date": "2023-08-24", "precision": "day", "category": "code", "open_weights": true, "params": "7B-34B", "note": "Llama 2 specialized for code with Python and Instruct variants", "source": "https://about.fb.com/news/2023/08/code-llama-ai-for-coding/" }, { "name": "HyperCLOVA X", "org": "Naver", "country": "South Korea", "date": "2023-08-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Korean-optimized LLM powering CLOVA X chatbot", "source": "https://techcrunch.com/2023/08/24/koreas-internet-giant-naver-unveils-generative-ai-services/" }, { "name": "Nougat", "org": "Meta", "country": "USA", "date": "2023-08-25", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "OCR model converting academic PDFs into markup", "source": "https://arxiv.org/abs/2308.13418" }, { "name": "Jais", "org": "G42", "country": "UAE", "date": "2023-08-30", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Open Arabic-centric LLM built with MBZUAI and Cerebras", "source": "https://arxiv.org/abs/2308.16149" }, { "name": "Swift", "org": "University of Zurich", "country": "Switzerland", "date": "2023-08-30", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Autonomous racing drone beat human world champions", "source": "https://www.nature.com/articles/s41586-023-06419-4" }, { "name": "CodeFuse-13B", "org": "Ant Group", "country": "China", "date": "2023-09-01", "precision": "month", "category": "code", "open_weights": true, "params": "13B", "note": "Ant Group's open multilingual code LLM", "source": "https://huggingface.co/api/models?author=codefuse-ai&sort=createdAt&direction=1&limit=15" }, { "name": "Nova-2", "org": "Deepgram", "country": "USA", "date": "2023-09-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Commercial speech-to-text model claiming 30% lower word error rate", "source": "https://deepgram.com/learn/nova-2-speech-to-text-api" }, { "name": "XTTS", "org": "Coqui", "country": "Germany", "date": "2023-09-01", "precision": "month", "category": "audio", "open_weights": true, "params": "", "note": "Open multilingual voice-cloning text-to-speech; XTTS-v2 followed weeks later", "source": "https://huggingface.co/coqui/XTTS-v1" }, { "name": "Baichuan 2", "org": "Baichuan", "country": "China", "date": "2023-09-06", "precision": "day", "category": "language", "open_weights": true, "params": "7B-13B", "note": "Second-generation open models trained on 2.6T tokens", "source": "https://github.com/baichuan-inc/Baichuan2" }, { "name": "Falcon 180B", "org": "TII", "country": "UAE", "date": "2023-09-06", "precision": "day", "category": "language", "open_weights": true, "params": "180B", "note": "Largest openly available LLM at release", "source": "https://huggingface.co/blog/falcon-180b" }, { "name": "Granite 13B", "org": "IBM", "country": "USA", "date": "2023-09-07", "precision": "day", "category": "language", "open_weights": false, "params": "13B", "note": "First Granite models for IBM watsonx", "source": "https://en.wikipedia.org/wiki/IBM_Granite" }, { "name": "Hunyuan", "org": "Tencent", "country": "China", "date": "2023-09-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Tencent's proprietary foundation model launched on Tencent Cloud", "source": "https://www.tencent.com/tencent-unveils-hunyuan-its-proprietary-large-foundation-model-on-tencent-cloud/" }, { "name": "Persimmon-8B", "org": "Adept", "country": "USA", "date": "2023-09-07", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Permissively licensed open model with 16K context", "source": "https://www.adept.ai/blog/persimmon-8b" }, { "name": "YandexGPT 2", "org": "Yandex", "country": "Russia", "date": "2023-09-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Improved Yandex model presented at Practical ML Conf", "source": "https://en.wikipedia.org/wiki/YandexGPT" }, { "name": "phi-1.5", "org": "Microsoft", "country": "USA", "date": "2023-09-11", "precision": "day", "category": "language", "open_weights": true, "params": "1.3B", "note": "Small model trained mostly on synthetic textbook data", "source": "https://arxiv.org/abs/2309.05463" }, { "name": "Stable Audio", "org": "Stability AI", "country": "UK", "date": "2023-09-13", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Latent diffusion music and sound effect generation", "source": "https://stability.ai/news-updates/stable-audio-using-ai-to-generate-music" }, { "name": "LINGO-1", "org": "Wayve", "country": "UK", "date": "2023-09-14", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Vision-language driving model narrating its decisions", "source": "https://wayve.ai/thinking/lingo-natural-language-autonomous-driving/" }, { "name": "AlphaMissense", "org": "Google DeepMind", "country": "UK", "date": "2023-09-19", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Classified 89% of 71 million possible missense variants", "source": "https://deepmind.google/blog/a-catalogue-of-genetic-mutations-to-help-pinpoint-the-cause-of-diseases/" }, { "name": "DALL·E 3", "org": "OpenAI", "country": "USA", "date": "2023-09-20", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Built natively with ChatGPT; far better prompt adherence", "source": "https://en.wikipedia.org/wiki/DALL-E" }, { "name": "InternLM-20B", "org": "Shanghai AI Lab", "country": "China", "date": "2023-09-20", "precision": "day", "category": "language", "open_weights": true, "params": "20B", "note": "Open 20B model released with base and chat versions", "source": "https://github.com/InternLM/InternLM" }, { "name": "GPT-4V", "org": "OpenAI", "country": "USA", "date": "2023-09-25", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "GPT-4 image understanding rolled out in ChatGPT", "source": "https://openai.com/index/gpt-4v-system-card/" }, { "name": "Qwen-14B", "org": "Alibaba", "country": "China", "date": "2023-09-25", "precision": "day", "category": "language", "open_weights": true, "params": "14B", "note": "Open 14B model trained on 3T tokens", "source": "https://github.com/QwenLM/Qwen" }, { "name": "InternLM-XComposer", "org": "Shanghai AI Lab", "country": "China", "date": "2023-09-26", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Vision-language model for interleaved text-image composition", "source": "https://arxiv.org/abs/2309.15112" }, { "name": "Emu", "org": "Meta", "country": "USA", "date": "2023-09-27", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Quality-tuned text-to-image model behind Meta AI image features", "source": "https://ai.meta.com/research/publications/emu-enhancing-image-generation-models-using-photogenic-needles-in-a-haystack/" }, { "name": "Mistral 7B", "org": "Mistral AI", "country": "France", "date": "2023-09-27", "precision": "day", "category": "language", "open_weights": true, "params": "7.3B", "note": "Apache 2.0 model beating Llama 2 13B, released via torrent", "source": "https://mistral.ai/news/announcing-mistral-7b/" }, { "name": "PLaMo-13B", "org": "Preferred Networks", "country": "Japan", "date": "2023-09-28", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Japanese-English open LLM from Preferred Networks", "source": "https://huggingface.co/pfnet/plamo-13b" }, { "name": "PixArt-α", "org": "Huawei", "country": "China", "date": "2023-09-30", "precision": "day", "category": "image", "open_weights": true, "params": "0.6B", "note": "Diffusion-transformer text-to-image at a fraction of SD training cost", "source": "https://arxiv.org/abs/2310.00426" }, { "name": "LLM-jp-13B", "org": "National Institute of Informatics", "country": "Japan", "date": "2023-10-01", "precision": "month", "category": "language", "open_weights": true, "params": "13B", "note": "Japanese academic consortium's open LLM", "source": "https://huggingface.co/llm-jp/llm-jp-13b-v1.0" }, { "name": "Zephyr-7B", "org": "Hugging Face", "country": "USA", "date": "2023-10-01", "precision": "month", "category": "language", "open_weights": true, "params": "7B", "note": "Mistral fine-tune aligned with distilled DPO", "source": "https://arxiv.org/abs/2310.16944" }, { "name": "RT-X", "org": "Google DeepMind", "country": "UK", "date": "2023-10-03", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Open X-Embodiment models trained on data from 22 robot types", "source": "https://deepmind.google/blog/scaling-up-learning-across-many-different-robot-types/" }, { "name": "CogVLM", "org": "Zhipu AI", "country": "China", "date": "2023-10-05", "precision": "day", "category": "language", "open_weights": true, "params": "17B", "note": "Visual-expert vision-language model", "source": "https://github.com/THUDM/CogVLM" }, { "name": "LLaVA-1.5", "org": "University of Wisconsin-Madison", "country": "USA", "date": "2023-10-05", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Simple upgrades made LLaVA state of the art among open VLMs", "source": "https://arxiv.org/abs/2310.03744" }, { "name": "Latent Consistency Models", "org": "Tsinghua University", "country": "China", "date": "2023-10-06", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Few-step, near-real-time latent diffusion image generation", "source": "https://arxiv.org/abs/2310.04378" }, { "name": "Kimi Chat", "org": "Moonshot AI", "country": "China", "date": "2023-10-09", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Long-context chatbot handling 200K Chinese characters", "source": "https://cn.chinadaily.com.cn/a/202310/10/WS652517eaa310d5acd87694a2.html" }, { "name": "RoseTTAFold All-Atom", "org": "University of Washington", "country": "USA", "date": "2023-10-09", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Predicts structures of proteins with ligands and nucleic acids", "source": "https://www.biorxiv.org/content/10.1101/2023.10.09.561603v1" }, { "name": "Firefly Design Model", "org": "Adobe", "country": "USA", "date": "2023-10-10", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Generates editable template designs for Adobe Express", "source": "https://news.adobe.com/news/news-details/2023/adobe-releases-next-generation-of-firefly-models" }, { "name": "Firefly Image 2", "org": "Adobe", "country": "USA", "date": "2023-10-10", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Second Firefly image model, announced at Adobe MAX", "source": "https://news.adobe.com/news/news-details/2023/adobe-releases-next-generation-of-firefly-models" }, { "name": "Firefly Vector Model", "org": "Adobe", "country": "USA", "date": "2023-10-10", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Billed as the first generative AI model for vector graphics", "source": "https://news.adobe.com/news/news-details/2023/adobe-releases-next-generation-of-firefly-models" }, { "name": "Ferret", "org": "Apple", "country": "USA", "date": "2023-10-11", "precision": "day", "category": "language", "open_weights": true, "params": "7B-13B", "note": "Apple's refer-and-ground multimodal LLM", "source": "https://arxiv.org/abs/2310.07704" }, { "name": "Aquila2", "org": "BAAI", "country": "China", "date": "2023-10-12", "precision": "day", "category": "language", "open_weights": true, "params": "7B-34B", "note": "Open bilingual Aquila2 7B/34B models", "source": "https://github.com/FlagAI-Open/Aquila2" }, { "name": "Llemma", "org": "EleutherAI", "country": "USA", "date": "2023-10-16", "precision": "day", "category": "science", "open_weights": true, "params": "7B-34B", "note": "Open language models for mathematics", "source": "https://arxiv.org/abs/2310.10631" }, { "name": "BitNet", "org": "Microsoft", "country": "USA", "date": "2023-10-17", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "1-bit Transformer architecture for LLMs", "source": "https://arxiv.org/abs/2310.11453" }, { "name": "ERNIE 4.0", "org": "Baidu", "country": "China", "date": "2023-10-17", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Baidu claimed parity with GPT-4", "source": "https://www.prnewswire.com/news-releases/baidu-launches-ernie-4-0-foundation-model-leading-a-new-wave-of-ai-native-applications-301958681.html" }, { "name": "Fuyu-8B", "org": "Adept", "country": "USA", "date": "2023-10-17", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Decoder-only multimodal model without a separate image encoder", "source": "https://www.adept.ai/blog/fuyu-8b" }, { "name": "Eureka", "org": "NVIDIA", "country": "USA", "date": "2023-10-19", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "GPT-4-written reward functions taught a robot hand pen spinning", "source": "https://arxiv.org/abs/2310.12931" }, { "name": "SparkDesk V3.0", "org": "iFlytek", "country": "China", "date": "2023-10-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "iFlytek claimed it outperformed ChatGPT in Chinese", "source": "https://www.scmp.com/tech/big-tech/article/3239031/iflytek-says-its-large-language-model-outperforms-chatgpt-chinese-ai-firm-vows-counter-us-chip-curbs" }, { "name": "jina-embeddings-v2", "org": "Jina AI", "country": "Germany", "date": "2023-10-25", "precision": "day", "category": "language", "open_weights": true, "params": "137M", "note": "First open-source text embedding model with 8K context", "source": "https://jina.ai/news/jina-ai-launches-worlds-first-open-source-8k-text-embedding-rivaling-openai/" }, { "name": "ChatGLM3", "org": "Zhipu AI", "country": "China", "date": "2023-10-27", "precision": "day", "category": "language", "open_weights": true, "params": "6B", "note": "Third-gen ChatGLM with tool calling and code execution", "source": "https://huggingface.co/zai-org/chatglm3-6b" }, { "name": "Skywork-13B", "org": "Kunlun Tech", "country": "China", "date": "2023-10-30", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Open model released with 150B-token Chinese corpus", "source": "https://arxiv.org/abs/2310.19341" }, { "name": "AlphaFold (next generation)", "org": "Google DeepMind", "country": "UK", "date": "2023-10-31", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Preview of AlphaFold 3: ligands, nucleic acids, nearly all PDB", "source": "https://deepmind.google/blog/a-glimpse-of-the-next-generation-of-alphafold/" }, { "name": "Tongyi Qianwen 2.0", "org": "Alibaba", "country": "China", "date": "2023-10-31", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Upgraded flagship with hundreds of billions of parameters", "source": "https://www.alibabacloud.com/blog/alibaba-cloud-launches-tongyi-qianwen-2-0-and-industry-specific-models-to-support-customers-reap-benefits-of-generative-ai_600526" }, { "name": "BlueLM", "org": "vivo", "country": "China", "date": "2023-11-01", "precision": "day", "category": "language", "open_weights": true, "params": "7B-175B", "note": "Smartphone-maker LLM family; 7B open-sourced", "source": "https://github.com/vivo-ai-lab/BlueLM" }, { "name": "Distil-Whisper", "org": "Hugging Face", "country": "USA", "date": "2023-11-01", "precision": "day", "category": "audio", "open_weights": true, "params": "756M", "note": "Distilled Whisper roughly six times faster for English", "source": "https://arxiv.org/abs/2311.00430" }, { "name": "OpenChat 3.5", "org": "Tsinghua University", "country": "China", "date": "2023-11-01", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "C-RLFT fine-tune of Mistral 7B rivaling ChatGPT on benchmarks", "source": "https://huggingface.co/openchat/openchat_3.5" }, { "name": "tsuzumi", "org": "NTT", "country": "Japan", "date": "2023-11-01", "precision": "day", "category": "language", "open_weights": false, "params": "0.6B-7B", "note": "Lightweight Japanese LLM from NTT", "source": "https://group.ntt/en/magazine/blog/tsuzumi/" }, { "name": "DeepSeek Coder", "org": "DeepSeek", "country": "China", "date": "2023-11-02", "precision": "day", "category": "code", "open_weights": true, "params": "1.3B-33B", "note": "DeepSeek's first release; top open code model", "source": "https://github.com/deepseek-ai/DeepSeek-Coder" }, { "name": "Embed v3", "org": "Cohere", "country": "Canada", "date": "2023-11-02", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Embedding model ranking document quality for RAG", "source": "https://cohere.com/blog/introducing-embed-v3" }, { "name": "Yi-34B", "org": "01.AI", "country": "China", "date": "2023-11-02", "precision": "day", "category": "language", "open_weights": true, "params": "6B-34B", "note": "Topped Hugging Face leaderboard for pretrained base models", "source": "https://github.com/01-ai/Yi" }, { "name": "Grok-1", "org": "xAI", "country": "USA", "date": "2023-11-03", "precision": "day", "category": "language", "open_weights": true, "params": "314B", "note": "xAI's debut model powering Grok on X; weights opened March 2024", "source": "https://x.ai/news/grok" }, { "name": "GPT-4 Turbo", "org": "OpenAI", "country": "USA", "date": "2023-11-06", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "128K-context, cheaper GPT-4 with vision announced at DevDay", "source": "https://openai.com/index/new-models-and-developer-products-announced-at-devday/" }, { "name": "TTS-1", "org": "OpenAI", "country": "USA", "date": "2023-11-06", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Text-to-speech API with six voices, plus TTS-1 HD variant", "source": "https://openai.com/index/new-models-and-developer-products-announced-at-devday/" }, { "name": "Whisper large-v3", "org": "OpenAI", "country": "USA", "date": "2023-11-06", "precision": "day", "category": "audio", "open_weights": true, "params": "1.55B", "note": "Improved open multilingual speech recognition", "source": "https://huggingface.co/openai/whisper-large-v3" }, { "name": "Samsung Gauss", "org": "Samsung", "country": "South Korea", "date": "2023-11-08", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Samsung's language, code and image generation models", "source": "https://techcrunch.com/2023/11/08/samsung-unveils-chatgpt-alternative-samsung-gauss-that-can-generate-text-code-and-images/" }, { "name": "Jais 30B", "org": "G42", "country": "UAE", "date": "2023-11-09", "precision": "day", "category": "language", "open_weights": true, "params": "30B", "note": "Scaled-up Arabic LLM from G42's Core42", "source": "https://www.prnewswire.com/ae/news-releases/core42-sets-new-benchmark-for-arabic-large-language-models-with-the-release-of-jais-30b-301983300.html" }, { "name": "Florence-2", "org": "Microsoft", "country": "USA", "date": "2023-11-10", "precision": "day", "category": "vision", "open_weights": true, "params": "0.23B-0.77B", "note": "Unified prompt-based model for many vision tasks", "source": "https://arxiv.org/abs/2311.06242" }, { "name": "NeuralGCM", "org": "Google", "country": "USA", "date": "2023-11-13", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Hybrid physics-ML general circulation model for weather and climate", "source": "https://arxiv.org/abs/2311.07222" }, { "name": "Qwen-Audio", "org": "Alibaba", "country": "China", "date": "2023-11-14", "precision": "day", "category": "audio", "open_weights": true, "params": "8B", "note": "Universal audio-language understanding model", "source": "https://arxiv.org/abs/2311.07919" }, { "name": "Nemotron-3 8B", "org": "NVIDIA", "country": "USA", "date": "2023-11-15", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "NVIDIA's enterprise-ready open LLM family", "source": "https://developer.nvidia.com/blog/nvidia-ai-foundation-models-build-custom-enterprise-chatbots-and-co-pilots-with-production-ready-llms/" }, { "name": "Phi-2", "org": "Microsoft", "country": "USA", "date": "2023-11-15", "precision": "day", "category": "language", "open_weights": true, "params": "2.7B", "note": "Announced at Ignite; small model matching models 25x larger", "source": "https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/" }, { "name": "AndesGPT", "org": "OPPO", "country": "China", "date": "2023-11-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "OPPO's LLM for its ColorOS phone assistant", "source": "https://www.oppo.com/en/newsroom/press/2023-oppo-developers-conference-odc23/" }, { "name": "Emu Edit", "org": "Meta", "country": "USA", "date": "2023-11-16", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Instruction-based precise image editing", "source": "https://ai.meta.com/blog/emu-text-to-video-generation-image-editing-research/" }, { "name": "Emu Video", "org": "Meta", "country": "USA", "date": "2023-11-16", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Factorized text-to-image-to-video generation", "source": "https://ai.meta.com/blog/emu-text-to-video-generation-image-editing-research/" }, { "name": "Lyria", "org": "Google DeepMind", "country": "UK", "date": "2023-11-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music model with vocals powering YouTube Dream Track", "source": "https://deepmind.google/blog/transforming-the-future-of-music-creation/" }, { "name": "Tülu 2", "org": "Allen Institute for AI", "country": "USA", "date": "2023-11-17", "precision": "day", "category": "language", "open_weights": true, "params": "7B-70B", "note": "DPO-tuned Llama 2 suite with fully open recipe", "source": "https://arxiv.org/abs/2311.10702" }, { "name": "Orca 2", "org": "Microsoft", "country": "USA", "date": "2023-11-18", "precision": "day", "category": "language", "open_weights": true, "params": "7B-13B", "note": "Small models taught multiple reasoning strategies", "source": "https://arxiv.org/abs/2311.11045" }, { "name": "Claude 2.1", "org": "Anthropic", "country": "USA", "date": "2023-11-21", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "200K context, halved hallucinations, tool use beta", "source": "https://www.anthropic.com/news/claude-2-1" }, { "name": "Stable Video Diffusion", "org": "Stability AI", "country": "UK", "date": "2023-11-21", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Stability AI's first open video generation model", "source": "https://stability.ai/news-updates/stable-video-diffusion-open-ai-video-model" }, { "name": "Inflection-2", "org": "Inflection AI", "country": "USA", "date": "2023-11-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Claimed second most capable LLM in the world at launch", "source": "http://web.archive.org/web/20231220170219/https://inflection.ai/inflection-2" }, { "name": "Kandinsky 3.0", "org": "Sber", "country": "Russia", "date": "2023-11-22", "precision": "day", "category": "image", "open_weights": true, "params": "11.9B", "note": "Russian open text-to-image diffusion model", "source": "https://github.com/ai-forever/Kandinsky-3" }, { "name": "Starling-7B", "org": "UC Berkeley", "country": "USA", "date": "2023-11-25", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "RLAIF-trained with Nectar dataset; top open 7B on MT-Bench", "source": "https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha" }, { "name": "MagicAnimate", "org": "ByteDance", "country": "China", "date": "2023-11-27", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Viral diffusion-based human image animation, built with NUS", "source": "https://arxiv.org/abs/2311.16498" }, { "name": "Meditron", "org": "EPFL", "country": "Switzerland", "date": "2023-11-27", "precision": "day", "category": "science", "open_weights": true, "params": "7B-70B", "note": "Open medical LLM adapted from Llama 2", "source": "https://arxiv.org/abs/2311.16079" }, { "name": "Yuan 2.0", "org": "Inspur", "country": "China", "date": "2023-11-27", "precision": "day", "category": "language", "open_weights": true, "params": "2B-102B", "note": "Open LLM with localized filtering-based attention", "source": "https://arxiv.org/abs/2311.15786" }, { "name": "Animate Anyone", "org": "Alibaba", "country": "China", "date": "2023-11-28", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Viral consistent character-animation image-to-video method", "source": "https://arxiv.org/abs/2311.17117" }, { "name": "Pika 1.0", "org": "Pika", "country": "USA", "date": "2023-11-28", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Consumer text-to-video and video editing launch", "source": "https://techcrunch.com/2023/11/28/pika-labs-which-is-building-ai-tools-to-generate-and-edit-videos-raises-55m/" }, { "name": "SDXL Turbo", "org": "Stability AI", "country": "UK", "date": "2023-11-28", "precision": "day", "category": "image", "open_weights": true, "params": "3.5B", "note": "Real-time single-step text-to-image via adversarial distillation", "source": "https://stability.ai/news-updates/stability-ai-sdxl-turbo" }, { "name": "Titan Image Generator", "org": "Amazon", "country": "USA", "date": "2023-11-28", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Invisibly watermarked text-to-image model, previewed on Bedrock", "source": "https://aws.amazon.com/about-aws/whats-new/2023/11/amazon-titan-image-generator-model-bedrock-preview/" }, { "name": "DeepSeek LLM", "org": "DeepSeek", "country": "China", "date": "2023-11-29", "precision": "day", "category": "language", "open_weights": true, "params": "7B-67B", "note": "DeepSeek's first general LLM, trained on 2T tokens", "source": "https://github.com/deepseek-ai/DeepSeek-LLM" }, { "name": "GNoME", "org": "Google DeepMind", "country": "UK", "date": "2023-11-29", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Predicted 2.2 million new crystals, 380,000 stable", "source": "https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/" }, { "name": "pplx-70b-online", "org": "Perplexity", "country": "USA", "date": "2023-11-29", "precision": "day", "category": "language", "open_weights": false, "params": "70B", "note": "Online LLMs with live web access via API", "source": "http://web.archive.org/web/20231229124455/https://blog.perplexity.ai/blog/introducing-pplx-online-llms" }, { "name": "Audiobox", "org": "Meta", "country": "USA", "date": "2023-11-30", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Voice and sound generation from natural-language prompts", "source": "https://ai.meta.com/blog/audiobox-generating-audio-voice-natural-language-prompts/" }, { "name": "Qwen-72B", "org": "Alibaba", "country": "China", "date": "2023-11-30", "precision": "day", "category": "language", "open_weights": true, "params": "72B", "note": "Largest open Qwen with 32K context, plus Qwen-1.8B", "source": "https://huggingface.co/Qwen/Qwen-72B" }, { "name": "Seamless", "org": "Meta", "country": "USA", "date": "2023-11-30", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Expressive, real-time streaming speech translation suite", "source": "https://ai.meta.com/blog/seamless-communication/" }, { "name": "Mamba", "org": "Carnegie Mellon University", "country": "USA", "date": "2023-12-01", "precision": "day", "category": "language", "open_weights": true, "params": "130M-2.8B", "note": "Selective state-space architecture rivaling Transformers", "source": "https://arxiv.org/abs/2312.00752" }, { "name": "AlphaCode 2", "org": "Google DeepMind", "country": "UK", "date": "2023-12-06", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Gemini-powered competitive programmer at 85th percentile", "source": "https://storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf" }, { "name": "Gemini 1.0 Nano", "org": "Google DeepMind", "country": "UK", "date": "2023-12-06", "precision": "day", "category": "multimodal", "open_weights": false, "params": "1.8B-3.25B", "note": "On-device Gemini debuting on Pixel 8 Pro", "source": "https://blog.google/technology/ai/google-gemini-ai" }, { "name": "Gemini 1.0 Pro", "org": "Google DeepMind", "country": "UK", "date": "2023-12-06", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Mid-size natively multimodal Gemini; powered Bard at launch", "source": "https://blog.google/technology/ai/google-gemini-ai" }, { "name": "Gemini 1.0 Ultra", "org": "Google DeepMind", "country": "UK", "date": "2023-12-06", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "First model reported to beat human experts on MMLU", "source": "https://blog.google/technology/ai/google-gemini-ai" }, { "name": "MatterGen", "org": "Microsoft", "country": "USA", "date": "2023-12-06", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Diffusion model generating novel inorganic materials", "source": "https://arxiv.org/abs/2312.03687" }, { "name": "Playground v2", "org": "Playground", "country": "USA", "date": "2023-12-06", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open SDXL-architecture model with strong aesthetic preference scores", "source": "https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic" }, { "name": "Llama Guard", "org": "Meta", "country": "USA", "date": "2023-12-07", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Input-output safety classifier launched with Purple Llama", "source": "https://arxiv.org/abs/2312.06674" }, { "name": "StableLM Zephyr 3B", "org": "Stability AI", "country": "UK", "date": "2023-12-07", "precision": "day", "category": "language", "open_weights": true, "params": "3B", "note": "Small chat model aimed at edge devices", "source": "https://stability.ai/news-updates/stablelm-zephyr-3b-stability-llm" }, { "name": "Mixtral 8x7B", "org": "Mistral AI", "country": "France", "date": "2023-12-08", "precision": "day", "category": "language", "open_weights": true, "params": "46.7B (12.9B active)", "note": "Sparse MoE matching GPT-3.5, dropped via torrent link", "source": "https://mistral.ai/news/mixtral-of-experts/" }, { "name": "StripedHyena-7B", "org": "Together", "country": "USA", "date": "2023-12-08", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Hybrid attention-convolution architecture competitive with Transformers", "source": "https://www.together.ai/blog/stripedhyena-7b" }, { "name": "Mistral Embed", "org": "Mistral AI", "country": "France", "date": "2023-12-11", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Mistral's first text-embedding model, launched with La Plateforme", "source": "https://mistral.ai/news/la-plateforme/" }, { "name": "Mistral Medium", "org": "Mistral AI", "country": "France", "date": "2023-12-11", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Mistral's first API-only flagship on La Plateforme", "source": "https://mistral.ai/news/la-plateforme/" }, { "name": "W.A.L.T", "org": "Google", "country": "USA", "date": "2023-12-11", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Transformer latent diffusion for photorealistic video, with Stanford", "source": "https://arxiv.org/abs/2312.06662" }, { "name": "Imagen 2", "org": "Google DeepMind", "country": "UK", "date": "2023-12-13", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Better photorealism plus multilingual text and logo rendering", "source": "https://deepmind.google/technologies/imagen-2/" }, { "name": "MedLM", "org": "Google", "country": "USA", "date": "2023-12-13", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Healthcare-tuned model family built on Med-PaLM 2", "source": "https://cloud.google.com/blog/topics/healthcare-life-sciences/introducing-medlm-for-the-healthcare-industry" }, { "name": "Octo", "org": "UC Berkeley", "country": "USA", "date": "2023-12-13", "precision": "day", "category": "robotics", "open_weights": true, "params": "93M", "note": "Open generalist robot policy trained on Open X-Embodiment", "source": "https://huggingface.co/api/models?author=rail-berkeley&sort=createdAt&direction=1&limit=20" }, { "name": "OpenHathi", "org": "Sarvam AI", "country": "India", "date": "2023-12-13", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Hindi-English bilingual open model built on Llama 2", "source": "https://huggingface.co/api/models?author=sarvamai&sort=createdAt&direction=1&limit=20" }, { "name": "SOLAR 10.7B", "org": "Upstage", "country": "South Korea", "date": "2023-12-13", "precision": "day", "category": "language", "open_weights": true, "params": "10.7B", "note": "Depth-upscaled model that topped the Open LLM Leaderboard", "source": "https://huggingface.co/upstage/SOLAR-10.7B-v1.0" }, { "name": "Stable Zero123", "org": "Stability AI", "country": "UK", "date": "2023-12-13", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Single-image to 3D object generation", "source": "https://stability.ai/news-updates/stable-zero123-3d-generation" }, { "name": "CogAgent", "org": "Zhipu AI", "country": "China", "date": "2023-12-14", "precision": "day", "category": "agent", "open_weights": true, "params": "18B", "note": "Visual language model that understands and operates GUIs", "source": "https://arxiv.org/abs/2312.08914" }, { "name": "FunSearch", "org": "Google DeepMind", "country": "UK", "date": "2023-12-14", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "LLM-driven program search found new cap set constructions", "source": "https://deepmind.google/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/" }, { "name": "Krutrim", "org": "Ola Krutrim", "country": "India", "date": "2023-12-15", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Ola's LLM billed as India's own AI", "source": "https://www.businesstoday.in/technology/news/story/indias-own-ai-krutrim-unveiled-by-ola-ceo-bhavish-aggarwal-check-availability-other-details-409604-2023-12-15" }, { "name": "Suno v2", "org": "Suno", "country": "USA", "date": "2023-12-19", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Full songs with vocals reach mass audience via Copilot and web app", "source": "https://en.wikipedia.org/wiki/Suno_AI" }, { "name": "VideoPoet", "org": "Google", "country": "USA", "date": "2023-12-19", "precision": "day", "category": "video", "open_weights": false, "params": "8B", "note": "LLM-based zero-shot video and audio generation", "source": "https://research.google/blog/videopoet-a-large-language-model-for-zero-shot-video-generation/" }, { "name": "Emu2", "org": "BAAI", "country": "China", "date": "2023-12-20", "precision": "day", "category": "multimodal", "open_weights": true, "params": "37B", "note": "Generative multimodal model with in-context learning", "source": "https://arxiv.org/abs/2312.13286" }, { "name": "InternVL", "org": "Shanghai AI Lab", "country": "China", "date": "2023-12-21", "precision": "day", "category": "vision", "open_weights": true, "params": "14B", "note": "Scaled 6B vision encoder aligned with LLMs", "source": "https://arxiv.org/abs/2312.14238" }, { "name": "Midjourney V6", "org": "Midjourney", "country": "USA", "date": "2023-12-21", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "In-image text and overhauled prompt understanding", "source": "https://venturebeat.com/ai/midjourney-v6-is-here-with-in-image-text-and-completely-overhauled-prompting/" }, { "name": "GenCast", "org": "Google DeepMind", "country": "UK", "date": "2023-12-25", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Diffusion-based ensemble medium-range weather forecasting", "source": "https://arxiv.org/abs/2312.15796" }, { "name": "KARAKURI LM 70B", "org": "KARAKURI", "country": "Japan", "date": "2024-01-01", "precision": "month", "category": "language", "open_weights": true, "params": "70B", "note": "Japanese-tuned open model built on Llama 2 70B", "source": "https://huggingface.co/karakuri-ai/karakuri-lm-70b-v0.1" }, { "name": "SFR-Embedding-Mistral", "org": "Salesforce", "country": "USA", "date": "2024-01-01", "precision": "month", "category": "language", "open_weights": true, "params": "7B", "note": "Mistral-based text embedding model that topped the MTEB leaderboard", "source": "https://huggingface.co/Salesforce/SFR-Embedding-Mistral" }, { "name": "AutoRT", "org": "Google DeepMind", "country": "UK", "date": "2024-01-04", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "LLM-orchestrated robot fleet data collection guided by a robot constitution", "source": "https://deepmind.google/discover/blog/shaping-the-future-of-advanced-robotics/" }, { "name": "Mobile ALOHA", "org": "Stanford University", "country": "USA", "date": "2024-01-04", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Low-cost bimanual mobile manipulator learning household tasks by imitation", "source": "https://arxiv.org/abs/2401.02117" }, { "name": "MagicVideo-V2", "org": "ByteDance", "country": "China", "date": "2024-01-09", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Multi-stage high-aesthetic text-to-video pipeline from ByteDance", "source": "https://arxiv.org/abs/2401.04468" }, { "name": "MAGNeT", "org": "Meta", "country": "USA", "date": "2024-01-09", "precision": "day", "category": "audio", "open_weights": true, "params": "1.5B", "note": "Non-autoregressive masked text-to-music and audio generator in AudioCraft", "source": "https://arxiv.org/abs/2401.04577" }, { "name": "DeepSeekMoE", "org": "DeepSeek", "country": "China", "date": "2024-01-11", "precision": "day", "category": "language", "open_weights": true, "params": "16.4B (2.8B active)", "note": "Fine-grained expert MoE design later scaled in DeepSeek-V2 and V3", "source": "https://arxiv.org/abs/2401.06066" }, { "name": "AMIE", "org": "Google", "country": "USA", "date": "2024-01-12", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Diagnostic dialogue AI outperforming physicians in simulated consultations", "source": "https://research.google/blog/amie-a-research-ai-system-for-diagnostic-medical-reasoning-and-conversations/" }, { "name": "InstantID", "org": "InstantX", "country": "China", "date": "2024-01-15", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Zero-shot identity-preserving image generation from a single face photo", "source": "https://arxiv.org/abs/2401.07519" }, { "name": "abab6", "org": "MiniMax", "country": "China", "date": "2024-01-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "MiniMax flagship billed as China's first MoE large language model", "source": "https://www.aibase.com/news/4893" }, { "name": "AIM (Autoregressive Image Models)", "org": "Apple", "country": "USA", "date": "2024-01-16", "precision": "day", "category": "vision", "open_weights": true, "params": "up to 7B", "note": "Vision encoders pre-trained autoregressively, scaling like LLMs", "source": "https://arxiv.org/abs/2401.08541" }, { "name": "GLM-4", "org": "Zhipu AI", "country": "China", "date": "2024-01-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Zhipu flagship claiming near GPT-4 parity with 128K context", "source": "https://www.maginative.com/article/zhipu-ai-unveils-next-gen-foundation-model-glm-4-claims-performance-comparable-to-gpt-4/" }, { "name": "Stable Code 3B", "org": "Stability AI", "country": "UK", "date": "2024-01-16", "precision": "day", "category": "code", "open_weights": true, "params": "3B", "note": "Laptop-sized code completion model matching CodeLlama 7B", "source": "https://stability.ai/news-updates/stable-code-2024-llm-code-completion-release" }, { "name": "AlphaGeometry", "org": "Google DeepMind", "country": "UK", "date": "2024-01-17", "precision": "day", "category": "science", "open_weights": true, "params": "151M", "note": "Neuro-symbolic system solving olympiad geometry near gold-medalist level", "source": "https://deepmind.google/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/" }, { "name": "DeciCoder-6B", "org": "Deci", "country": "Israel", "date": "2024-01-17", "precision": "day", "category": "code", "open_weights": true, "params": "6B", "note": "Efficient open code model designed with neural architecture search", "source": "https://huggingface.co/Deci/DeciCoder-6B" }, { "name": "InternLM2", "org": "Shanghai AI Lab", "country": "China", "date": "2024-01-17", "precision": "day", "category": "language", "open_weights": true, "params": "7B, 20B", "note": "Open bilingual LLMs with 200K context, co-developed with SenseTime", "source": "https://github.com/InternLM/InternLM" }, { "name": "VideoCrafter2", "org": "Tencent", "country": "China", "date": "2024-01-17", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open text-to-video diffusion trained around data limitations", "source": "https://arxiv.org/abs/2401.09047" }, { "name": "Vision Mamba (Vim)", "org": "Huazhong University of Science and Technology", "country": "China", "date": "2024-01-17", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Bidirectional state-space backbone challenging vision transformers", "source": "https://arxiv.org/abs/2401.09417" }, { "name": "Depth Anything", "org": "ByteDance", "country": "China", "date": "2024-01-19", "precision": "day", "category": "vision", "open_weights": true, "params": "25M-335M", "note": "Monocular depth foundation model trained on large-scale unlabeled images", "source": "https://arxiv.org/abs/2401.10891" }, { "name": "Stable LM 2 1.6B", "org": "Stability AI", "country": "UK", "date": "2024-01-19", "precision": "day", "category": "language", "open_weights": true, "params": "1.6B", "note": "Small multilingual LLM trained on 2T tokens in seven languages", "source": "https://stability.ai/news-updates/introducing-stable-lm-2" }, { "name": "Lumiere", "org": "Google", "country": "USA", "date": "2024-01-23", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Space-time U-Net generating an entire video clip in one pass", "source": "https://arxiv.org/abs/2401.12945" }, { "name": "voyage-code-2", "org": "Voyage AI", "country": "USA", "date": "2024-01-23", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Embedding model specialized for code retrieval in RAG", "source": "https://blog.voyageai.com/2024/01/23/voyage-code-2-elevate-your-code-retrieval/" }, { "name": "Yi-VL", "org": "01.AI", "country": "China", "date": "2024-01-23", "precision": "day", "category": "language", "open_weights": true, "params": "6B, 34B", "note": "Open Yi vision-language models; 34B topped MMMU among open models", "source": "https://github.com/01-ai/Yi" }, { "name": "Fuyu-Heavy", "org": "Adept", "country": "USA", "date": "2024-01-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Adept's large multimodal model aimed at UI understanding for agents", "source": "https://www.adept.ai/blog/adept-fuyu-heavy" }, { "name": "Qwen-VL-Max", "org": "Alibaba", "country": "China", "date": "2024-01-25", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Alibaba's top vision-language API model, pitched against GPT-4V", "source": "https://qwenlm.github.io/blog/qwen-vl/" }, { "name": "text-embedding-3-large", "org": "OpenAI", "country": "USA", "date": "2024-01-25", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "OpenAI's strongest embedder, 3072 dims with Matryoshka shortening", "source": "https://openai.com/index/new-embedding-models-and-api-updates/" }, { "name": "text-embedding-3-small", "org": "OpenAI", "country": "USA", "date": "2024-01-25", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Cheaper, stronger successor to ada-002 embeddings", "source": "https://openai.com/index/new-embedding-models-and-api-updates/" }, { "name": "Baichuan 3", "org": "Baichuan", "country": "China", "date": "2024-01-29", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Baichuan's third-generation proprietary flagship LLM", "source": "https://www.jiqizhixin.com/articles/2024-01-29-12" }, { "name": "Code Llama 70B", "org": "Meta", "country": "USA", "date": "2024-01-29", "precision": "day", "category": "code", "open_weights": true, "params": "70B", "note": "Largest Code Llama; Instruct variant scored 67.8 on HumanEval", "source": "https://siliconangle.com/2024/01/29/meta-releases-more-powerful-code-llama-70b-model-writing-software/" }, { "name": "Eagle 7B (RWKV-v5)", "org": "RWKV Foundation", "country": "USA", "date": "2024-01-29", "precision": "day", "category": "language", "open_weights": true, "params": "7.52B", "note": "Attention-free RNN LLM trained on 1T tokens across 100+ languages", "source": "https://blog.rwkv.com/p/eagle-7b-soaring-past-transformers" }, { "name": "InternLM-XComposer2", "org": "Shanghai AI Lab", "country": "China", "date": "2024-01-29", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Open vision-language model for free-form text-image composition", "source": "https://arxiv.org/abs/2401.16420" }, { "name": "Niji V6", "org": "Midjourney", "country": "USA", "date": "2024-01-29", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Anime-tuned model built on the Midjourney V6 generation", "source": "https://x.com/midjourney/status/1752115495065755798" }, { "name": "BGE-M3", "org": "BAAI", "country": "China", "date": "2024-01-30", "precision": "day", "category": "language", "open_weights": true, "params": "568M", "note": "Multilingual, multi-granularity embedding model", "source": "https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/BGE_M3/BGE_M3.pdf" }, { "name": "iFlytek Spark V3.5", "org": "iFlytek", "country": "China", "date": "2024-01-30", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "First model trained on fully domestic Feixing-1 compute platform", "source": "https://huacheng.gz-cmc.com/pages/2024/01/30/SF11469802a2a3c880b2d24478b0ddfa.html" }, { "name": "LLaVA-NeXT (LLaVA-1.6)", "org": "University of Wisconsin-Madison", "country": "USA", "date": "2024-01-30", "precision": "day", "category": "language", "open_weights": true, "params": "7B-34B", "note": "Open multimodal model with higher resolution, better OCR and reasoning", "source": "https://llava-vl.github.io/blog/2024-01-30-llava-next/" }, { "name": "YOLO-World", "org": "Tencent", "country": "China", "date": "2024-01-30", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Real-time open-vocabulary object detection", "source": "https://arxiv.org/abs/2401.17270" }, { "name": "Canary-1B", "org": "NVIDIA", "country": "USA", "date": "2024-02-01", "precision": "month", "category": "audio", "open_weights": true, "params": "1B", "note": "Multilingual speech recognition and translation model from NeMo", "source": "https://huggingface.co/nvidia/canary-1b" }, { "name": "MetaVoice-1B", "org": "MetaVoice", "country": "UK", "date": "2024-02-01", "precision": "month", "category": "audio", "open_weights": true, "params": "1.2B", "note": "Open TTS model with zero-shot voice cloning", "source": "https://huggingface.co/metavoiceio/metavoice-1B-v0.1" }, { "name": "MiniCPM", "org": "ModelBest", "country": "China", "date": "2024-02-01", "precision": "day", "category": "language", "open_weights": true, "params": "2.4B", "note": "Edge-sized LLM claiming Mistral-7B-level performance; MiniCPM-V alongside", "source": "https://github.com/OpenBMB/MiniCPM" }, { "name": "Nomic Embed", "org": "Nomic AI", "country": "USA", "date": "2024-02-01", "precision": "month", "category": "language", "open_weights": true, "params": "137M", "note": "Fully open 8192-context text embedder with open training data", "source": "https://arxiv.org/abs/2402.01613" }, { "name": "OLMo", "org": "Allen Institute for AI", "country": "USA", "date": "2024-02-01", "precision": "day", "category": "language", "open_weights": true, "params": "1B, 7B", "note": "Fully open LLM with released data, code, checkpoints and logs", "source": "https://allenai.org/blog/olmo-open-language-model-87ccfc95f580" }, { "name": "Smaug-72B", "org": "Abacus.AI", "country": "USA", "date": "2024-02-01", "precision": "month", "category": "language", "open_weights": true, "params": "72B", "note": "First open model averaging above 80 on Hugging Face leaderboard", "source": "https://huggingface.co/abacusai/Smaug-72B-v0.1" }, { "name": "Stable Video Diffusion 1.1", "org": "Stability AI", "country": "UK", "date": "2024-02-01", "precision": "month", "category": "video", "open_weights": true, "params": "", "note": "Fine-tuned image-to-video model with improved consistency", "source": "https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt-1-1" }, { "name": "TimesFM", "org": "Google", "country": "USA", "date": "2024-02-02", "precision": "day", "category": "science", "open_weights": true, "params": "200M", "note": "Decoder-only foundation model for zero-shot time-series forecasting", "source": "https://research.google/blog/a-decoder-only-foundation-model-for-time-series-forecasting/" }, { "name": "Moirai", "org": "Salesforce", "country": "USA", "date": "2024-02-04", "precision": "day", "category": "science", "open_weights": true, "params": "14M-311M", "note": "Universal time-series forecasting foundation model", "source": "https://arxiv.org/abs/2402.02592" }, { "name": "Qwen1.5", "org": "Alibaba", "country": "China", "date": "2024-02-04", "precision": "day", "category": "language", "open_weights": true, "params": "0.5B-72B", "note": "Open Qwen family refresh spanning 0.5B to 72B with 32K context", "source": "https://qwenlm.github.io/blog/qwen1.5/" }, { "name": "DeepSeekMath", "org": "DeepSeek", "country": "China", "date": "2024-02-05", "precision": "day", "category": "science", "open_weights": true, "params": "7B", "note": "Math LLM that introduced GRPO, later central to DeepSeek-R1", "source": "https://arxiv.org/abs/2402.03300" }, { "name": "EVA-CLIP-18B", "org": "BAAI", "country": "China", "date": "2024-02-06", "precision": "day", "category": "vision", "open_weights": true, "params": "18B", "note": "Largest open CLIP model at release", "source": "https://arxiv.org/abs/2402.04252" }, { "name": "SenseNova 4.0", "org": "SenseTime", "country": "China", "date": "2024-02-06", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Model suite upgrade incl. SenseChat V4 and 10B SenseMirage V4", "source": "https://www.prnewswire.com/apac/news-releases/sensetime-unveils-sensenova-4-0--bringing-novel-ai-experience-302054548.html" }, { "name": "Grandmaster-level chess transformer", "org": "Google DeepMind", "country": "UK", "date": "2024-02-07", "precision": "day", "category": "games", "open_weights": true, "params": "270M", "note": "Transformer reaching grandmaster-level blitz play without explicit search", "source": "https://arxiv.org/abs/2402.04494" }, { "name": "LGM (Large Multi-View Gaussian Model)", "org": "Peking University", "country": "China", "date": "2024-02-07", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Fast text/image-to-3D via multi-view Gaussian splatting", "source": "https://arxiv.org/abs/2402.05054" }, { "name": "Spirit LM", "org": "Meta", "country": "USA", "date": "2024-02-08", "precision": "day", "category": "audio", "open_weights": true, "params": "7B", "note": "Interleaved spoken and written language model", "source": "https://arxiv.org/abs/2402.05755" }, { "name": "BASE TTS", "org": "Amazon", "country": "USA", "date": "2024-02-12", "precision": "day", "category": "audio", "open_weights": false, "params": "1B", "note": "Largest TTS model at publication, trained on 100K hours of speech", "source": "https://arxiv.org/abs/2402.08093" }, { "name": "Reka Edge", "org": "Reka", "country": "USA", "date": "2024-02-12", "precision": "day", "category": "multimodal", "open_weights": false, "params": "7B", "note": "Compact 7B multimodal model for local and on-device use", "source": "https://reka.ai/news/reka-flash-efficient-and-capable-multimodal-language-models" }, { "name": "Reka Flash", "org": "Reka", "country": "USA", "date": "2024-02-12", "precision": "day", "category": "multimodal", "open_weights": false, "params": "21B", "note": "21B natively multimodal model competitive with larger rivals", "source": "https://reka.ai/news/reka-flash-efficient-and-capable-multimodal-language-models" }, { "name": "Stable Cascade", "org": "Stability AI", "country": "UK", "date": "2024-02-12", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Wurstchen-based cascaded text-to-image with tiny latent space", "source": "https://stability.ai/news/introducing-stable-cascade" }, { "name": "Aya 101", "org": "Cohere", "country": "Canada", "date": "2024-02-13", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Cohere For AI open multilingual model covering 101 languages", "source": "https://cohere.com/research/aya" }, { "name": "Gemini 1.5 Pro", "org": "Google DeepMind", "country": "UK", "date": "2024-02-15", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "MoE multimodal model with a 1M-token context window", "source": "https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/" }, { "name": "Sora", "org": "OpenAI", "country": "USA", "date": "2024-02-15", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Minute-long text-to-video framed as a world simulator", "source": "https://openai.com/index/video-generation-models-as-world-simulators/" }, { "name": "Universal Manipulation Interface (UMI)", "org": "Stanford University", "country": "USA", "date": "2024-02-15", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Handheld-gripper data collection enabling in-the-wild robot policies", "source": "https://arxiv.org/abs/2402.10329" }, { "name": "V-JEPA", "org": "Meta", "country": "USA", "date": "2024-02-15", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "LeCun's joint-embedding predictive architecture learning from video", "source": "https://ai.meta.com/research/publications/revisiting-feature-prediction-for-learning-visual-representations-from-video/" }, { "name": "Orca-Math", "org": "Microsoft", "country": "USA", "date": "2024-02-16", "precision": "day", "category": "science", "open_weights": false, "params": "7B", "note": "Small Mistral-based model specialized for grade-school math", "source": "https://arxiv.org/abs/2402.14830" }, { "name": "VideoPrism", "org": "Google", "country": "USA", "date": "2024-02-20", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Foundational video encoder for general video understanding", "source": "https://arxiv.org/abs/2402.13217" }, { "name": "Gemma", "org": "Google DeepMind", "country": "UK", "date": "2024-02-21", "precision": "day", "category": "language", "open_weights": true, "params": "2B, 7B", "note": "Google's first open-weight LLMs derived from Gemini research", "source": "https://blog.google/technology/developers/gemma-open-models/" }, { "name": "SDXL-Lightning", "org": "ByteDance", "country": "China", "date": "2024-02-21", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Distilled SDXL producing 1024px images in one to eight steps", "source": "https://arxiv.org/abs/2402.13929" }, { "name": "YOLOv9", "org": "Academia Sinica", "country": "Taiwan", "date": "2024-02-21", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Programmable gradient information advances real-time detection", "source": "https://arxiv.org/abs/2402.13616" }, { "name": "MobileLLM", "org": "Meta", "country": "USA", "date": "2024-02-22", "precision": "day", "category": "language", "open_weights": true, "params": "125M-350M", "note": "Sub-billion-parameter LLM architecture for on-device use", "source": "https://arxiv.org/abs/2402.14905" }, { "name": "Snap Video", "org": "Snap", "country": "USA", "date": "2024-02-22", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Scaled spatiotemporal transformer for text-to-video", "source": "https://arxiv.org/abs/2402.14797" }, { "name": "Stable Diffusion 3", "org": "Stability AI", "country": "UK", "date": "2024-02-22", "precision": "day", "category": "image", "open_weights": false, "params": "800M-8B", "note": "Multimodal diffusion transformer with rectified flow; preview announced", "source": "https://stability.ai/news-updates/stable-diffusion-3" }, { "name": "Genie", "org": "Google DeepMind", "country": "UK", "date": "2024-02-23", "precision": "day", "category": "video", "open_weights": false, "params": "11B", "note": "Generative interactive world model learned from unlabeled game videos", "source": "https://arxiv.org/abs/2402.15391" }, { "name": "Mistral Large", "org": "Mistral AI", "country": "France", "date": "2024-02-26", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Mistral's flagship, launched on la Plateforme, Azure and Le Chat", "source": "https://mistral.ai/news/mistral-large/" }, { "name": "Mistral Small", "org": "Mistral AI", "country": "France", "date": "2024-02-26", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Low-latency proprietary model launched alongside Mistral Large", "source": "https://mistral.ai/news/mistral-large/" }, { "name": "Nemotron-4 15B", "org": "NVIDIA", "country": "USA", "date": "2024-02-26", "precision": "day", "category": "language", "open_weights": false, "params": "15B", "note": "Multilingual 15B model trained on 8T tokens", "source": "https://arxiv.org/abs/2402.16819" }, { "name": "BitNet b1.58", "org": "Microsoft", "country": "USA", "date": "2024-02-27", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Ternary 1.58-bit LLMs matching full-precision performance", "source": "https://arxiv.org/abs/2402.17764" }, { "name": "EMO (Emote Portrait Alive)", "org": "Alibaba", "country": "China", "date": "2024-02-27", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Audio-driven expressive talking and singing portrait videos", "source": "https://arxiv.org/abs/2402.17485" }, { "name": "Evo", "org": "Arc Institute", "country": "USA", "date": "2024-02-27", "precision": "day", "category": "science", "open_weights": true, "params": "7B", "note": "StripedHyena DNA foundation model generating genome-scale sequences", "source": "https://arcinstitute.org/news/blog/evo" }, { "name": "Palmyra Vision", "org": "Writer", "country": "USA", "date": "2024-02-27", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Writer's enterprise multimodal LLM that analyzes images", "source": "https://writer.com/blog/palmyra-vision/" }, { "name": "Playground v2.5", "org": "Playground", "country": "USA", "date": "2024-02-27", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open SDXL-architecture model focused on aesthetic quality", "source": "https://arxiv.org/abs/2402.17245" }, { "name": "Ideogram 1.0", "org": "Ideogram", "country": "Canada", "date": "2024-02-28", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Text-to-image model noted for accurate in-image typography", "source": "https://about.ideogram.ai/1.0" }, { "name": "StarCoder2", "org": "BigCode", "country": "USA", "date": "2024-02-28", "precision": "day", "category": "code", "open_weights": true, "params": "3B, 7B, 15B", "note": "Open code LLMs from Hugging Face, ServiceNow, NVIDIA on The Stack v2", "source": "https://huggingface.co/blog/starcoder2" }, { "name": "Griffin", "org": "Google DeepMind", "country": "UK", "date": "2024-02-29", "precision": "day", "category": "language", "open_weights": false, "params": "up to 14B", "note": "Hybrid gated linear recurrence plus local attention; Hawk sibling", "source": "https://arxiv.org/abs/2402.19427" }, { "name": "Aramco Metabrain AI", "org": "Aramco", "country": "Saudi Arabia", "date": "2024-03-04", "precision": "day", "category": "language", "open_weights": false, "params": "250B", "note": "Saudi Aramco's industrial generative AI model unveiled at LEAP", "source": "https://www.offshore-technology.com/news/saudi-aramco-unveils-industry-first-generative-ai-model/" }, { "name": "Claude 3 Haiku", "org": "Anthropic", "country": "USA", "date": "2024-03-04", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Fastest, cheapest Claude 3; announced March 4, released March 13", "source": "https://www.anthropic.com/news/claude-3-haiku" }, { "name": "Claude 3 Opus", "org": "Anthropic", "country": "USA", "date": "2024-03-04", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Most capable Claude 3 model, outscoring GPT-4 on many benchmarks", "source": "https://www.anthropic.com/news/claude-3-family" }, { "name": "Claude 3 Sonnet", "org": "Anthropic", "country": "USA", "date": "2024-03-04", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Balanced Claude 3 tier powering free claude.ai", "source": "https://www.anthropic.com/news/claude-3-family" }, { "name": "TripoSR", "org": "Stability AI", "country": "UK", "date": "2024-03-04", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Sub-second single-image 3D reconstruction, built with Tripo AI", "source": "https://stability.ai/news-updates/triposr-3d-generation" }, { "name": "NaturalSpeech 3", "org": "Microsoft", "country": "USA", "date": "2024-03-05", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Zero-shot TTS using factorized codec and diffusion", "source": "https://arxiv.org/abs/2403.03100" }, { "name": "Yi-9B", "org": "01.AI", "country": "China", "date": "2024-03-06", "precision": "day", "category": "language", "open_weights": true, "params": "9B", "note": "Depth-upscaled Yi model strengthened in code and math", "source": "https://github.com/01-ai/Yi" }, { "name": "Inflection-2.5", "org": "Inflection AI", "country": "USA", "date": "2024-03-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Upgraded model behind Pi, claimed near GPT-4 performance", "source": "https://inflection.ai/inflection-2-5" }, { "name": "PixArt-Σ", "org": "Huawei", "country": "China", "date": "2024-03-07", "precision": "day", "category": "image", "open_weights": true, "params": "0.6B", "note": "Noah's Ark DiT enabling direct 4K text-to-image generation", "source": "https://arxiv.org/abs/2403.04692" }, { "name": "CogView3", "org": "Zhipu AI", "country": "China", "date": "2024-03-08", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Relay-diffusion text-to-image model from Zhipu and Tsinghua", "source": "https://arxiv.org/abs/2403.05121" }, { "name": "DeepSeek-VL", "org": "DeepSeek", "country": "China", "date": "2024-03-08", "precision": "day", "category": "language", "open_weights": true, "params": "1.3B, 7B", "note": "Open vision-language models for real-world understanding", "source": "https://arxiv.org/abs/2403.05525" }, { "name": "mxbai-embed-large-v1", "org": "Mixedbread", "country": "Germany", "date": "2024-03-08", "precision": "day", "category": "language", "open_weights": true, "params": "335M", "note": "Open English embedding model from Berlin-based Mixedbread", "source": "https://www.mixedbread.com/blog/mxbai-embed-large-v1" }, { "name": "Command R", "org": "Cohere", "country": "Canada", "date": "2024-03-11", "precision": "day", "category": "language", "open_weights": true, "params": "35B", "note": "RAG and tool-use optimized model with open research weights", "source": "https://cohere.com/blog/command-r" }, { "name": "RFM-1", "org": "Covariant", "country": "USA", "date": "2024-03-11", "precision": "day", "category": "robotics", "open_weights": false, "params": "8B", "note": "Robotics foundation model trained on warehouse robot data", "source": "https://covariant.ai/insights/introducing-rfm-1-giving-robots-human-like-reasoning-capabilities/" }, { "name": "Chronos", "org": "Amazon", "country": "USA", "date": "2024-03-12", "precision": "day", "category": "science", "open_weights": true, "params": "8M-710M", "note": "Time-series forecasting models using language-model tokenization", "source": "https://arxiv.org/abs/2403.07815" }, { "name": "Devin", "org": "Cognition", "country": "USA", "date": "2024-03-12", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Billed as the first autonomous AI software engineer", "source": "https://www.cognition.ai/blog/introducing-devin" }, { "name": "Recraft V2", "org": "Recraft", "country": "UK", "date": "2024-03-13", "precision": "day", "category": "image", "open_weights": false, "params": "20B", "note": "20B image model built for designers, including vector output", "source": "https://www.recraft.ai/blog/recraft-20b" }, { "name": "SIMA", "org": "Google DeepMind", "country": "UK", "date": "2024-03-13", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Generalist agent following language instructions across 3D video games", "source": "https://deepmind.google/discover/blog/sima-generalist-ai-agent-for-3d-virtual-environments/" }, { "name": "VLOGGER", "org": "Google", "country": "USA", "date": "2024-03-13", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Audio-driven full-body avatar video from a single image", "source": "https://arxiv.org/abs/2403.08764" }, { "name": "MM1", "org": "Apple", "country": "USA", "date": "2024-03-14", "precision": "day", "category": "language", "open_weights": false, "params": "up to 30B", "note": "Apple's multimodal LLM pre-training study, dense and MoE", "source": "https://arxiv.org/abs/2403.09611" }, { "name": "Quiet-STaR", "org": "Stanford University", "country": "USA", "date": "2024-03-14", "precision": "day", "category": "language", "open_weights": false, "params": "7B", "note": "LMs learn to generate internal rationales before speaking", "source": "https://arxiv.org/abs/2403.09629" }, { "name": "Open-Sora 1.0", "org": "HPC-AI Tech", "country": "Singapore", "date": "2024-03-18", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Fully open-source Sora-style video generation pipeline", "source": "https://github.com/hpcaitech/Open-Sora" }, { "name": "Project GR00T", "org": "NVIDIA", "country": "USA", "date": "2024-03-18", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Foundation model project for general-purpose humanoid robots", "source": "https://nvidianews.nvidia.com/news/foundation-model-isaac-robotics-platform" }, { "name": "Stable Video 3D (SV3D)", "org": "Stability AI", "country": "UK", "date": "2024-03-18", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Orbital novel-view synthesis and 3D generation from a single image", "source": "https://stability.ai/news-updates/introducing-stable-video-3d" }, { "name": "AnimateDiff-Lightning", "org": "ByteDance", "country": "China", "date": "2024-03-19", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Few-step distilled text-to-video generation", "source": "https://arxiv.org/abs/2403.12706" }, { "name": "TacticAI", "org": "Google DeepMind", "country": "UK", "date": "2024-03-19", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "AI assistant suggesting football corner-kick tactics", "source": "https://deepmind.google/discover/blog/tacticai-ai-assistant-for-football-tactics/" }, { "name": "EvoLLM-JP", "org": "Sakana AI", "country": "Japan", "date": "2024-03-21", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Japanese math LLM created by evolutionary model merging", "source": "https://sakana.ai/evolutionary-model-merge/" }, { "name": "Rakuten AI 7B", "org": "Rakuten", "country": "Japan", "date": "2024-03-21", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Japanese-optimized open LLM based on Mistral 7B", "source": "https://global.rakuten.com/corp/news/press/2024/0321_01.html" }, { "name": "Suno v3", "org": "Suno", "country": "USA", "date": "2024-03-21", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Full songs with vocals from text, released to all users", "source": "https://suno.com/blog/v3" }, { "name": "Step-1", "org": "StepFun", "country": "China", "date": "2024-03-23", "precision": "day", "category": "language", "open_weights": false, "params": "100B-class", "note": "StepFun's debut hundred-billion-parameter language model", "source": "https://www.cls.cn/detail/1627876" }, { "name": "Step-1V", "org": "StepFun", "country": "China", "date": "2024-03-23", "precision": "day", "category": "language", "open_weights": false, "params": "100B-class", "note": "Hundred-billion-parameter multimodal model topping OpenCompass at launch", "source": "https://www.cls.cn/detail/1627876" }, { "name": "Step-2", "org": "StepFun", "country": "China", "date": "2024-03-23", "precision": "day", "category": "language", "open_weights": false, "params": "1T (MoE)", "note": "Trillion-parameter MoE LLM previewed to select partners", "source": "https://www.cls.cn/detail/1627876" }, { "name": "EVI (Empathic Voice Interface)", "org": "Hume AI", "country": "USA", "date": "2024-03-25", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Voice AI that detects and responds to vocal emotion", "source": "https://www.hume.ai/blog/series-b-evi-announcement" }, { "name": "Stable Code Instruct 3B", "org": "Stability AI", "country": "UK", "date": "2024-03-25", "precision": "day", "category": "code", "open_weights": true, "params": "3B", "note": "Instruction-tuned Stable Code for natural-language coding tasks", "source": "https://stability.ai/news-updates/introducing-stable-code-instruct-3b" }, { "name": "VoiceCraft", "org": "University of Texas at Austin", "country": "USA", "date": "2024-03-25", "precision": "day", "category": "audio", "open_weights": true, "params": "830M", "note": "Open codec language model for speech editing and zero-shot TTS", "source": "https://arxiv.org/abs/2403.16973" }, { "name": "AniPortrait", "org": "Tencent", "country": "China", "date": "2024-03-26", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Audio-driven photorealistic portrait animation", "source": "https://arxiv.org/abs/2403.17694" }, { "name": "DBRX", "org": "Databricks", "country": "USA", "date": "2024-03-27", "precision": "day", "category": "language", "open_weights": true, "params": "132B (36B active)", "note": "Fine-grained MoE open model beating Llama 2 70B and Mixtral", "source": "https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm" }, { "name": "Grok-1.5", "org": "xAI", "country": "USA", "date": "2024-03-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Improved reasoning and 128K context for Grok", "source": "https://x.ai/news/grok-1.5" }, { "name": "Jamba", "org": "AI21 Labs", "country": "Israel", "date": "2024-03-28", "precision": "day", "category": "language", "open_weights": true, "params": "52B (12B active)", "note": "First production-scale hybrid Mamba-Transformer MoE model", "source": "https://www.ai21.com/blog/announcing-jamba" }, { "name": "Qwen1.5-MoE-A2.7B", "org": "Alibaba", "country": "China", "date": "2024-03-28", "precision": "day", "category": "language", "open_weights": true, "params": "14.3B (2.7B active)", "note": "MoE matching 7B models with a third of activated parameters", "source": "https://qwenlm.github.io/blog/qwen-moe/" }, { "name": "YandexGPT 3 Pro", "org": "Yandex", "country": "Russia", "date": "2024-03-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Yandex's third-generation Russian-language flagship LLM", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Gecko", "org": "Google DeepMind", "country": "UK", "date": "2024-03-29", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Compact text embedding model distilled from LLMs", "source": "https://arxiv.org/abs/2403.20327" }, { "name": "ReALM", "org": "Apple", "country": "USA", "date": "2024-03-29", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Reference resolution including on-screen context", "source": "https://arxiv.org/abs/2403.20329" }, { "name": "Voice Engine", "org": "OpenAI", "country": "USA", "date": "2024-03-29", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Voice cloning from a 15-second sample; limited preview only", "source": "https://openai.com/blog/navigating-the-challenges-and-opportunities-of-synthetic-voices" }, { "name": "Bielik-7B-v0.1", "org": "SpeakLeash", "country": "Poland", "date": "2024-04-01", "precision": "month", "category": "language", "open_weights": true, "params": "7B", "note": "Polish open LLM from SpeakLeash and ACK Cyfronet", "source": "https://pl.wikipedia.org/wiki/Bielik_(model_językowy)" }, { "name": "Qwen1.5-32B", "org": "Alibaba", "country": "China", "date": "2024-04-02", "precision": "day", "category": "language", "open_weights": true, "params": "32B", "note": "Mid-size Qwen1.5 balancing capability and efficiency", "source": "https://qwenlm.github.io/blog/qwen1.5-32b/" }, { "name": "SWE-agent", "org": "Princeton University", "country": "USA", "date": "2024-04-02", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Open agent-computer interface letting LMs fix real GitHub issues", "source": "https://github.com/princeton-nlp/SWE-agent" }, { "name": "Stable Audio 2.0", "org": "Stability AI", "country": "UK", "date": "2024-04-03", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Full three-minute tracks plus audio-to-audio generation", "source": "https://stability.ai/news/stable-audio-2-0" }, { "name": "Universal-1", "org": "AssemblyAI", "country": "USA", "date": "2024-04-03", "precision": "day", "category": "audio", "open_weights": false, "params": "600M", "note": "Speech recognition model trained on 12.5M hours of multilingual audio", "source": "https://www.assemblyai.com/blog/announcing-universal-1-speech-recognition-model" }, { "name": "Command R+", "org": "Cohere", "country": "Canada", "date": "2024-04-04", "precision": "day", "category": "language", "open_weights": true, "params": "104B", "note": "Enterprise RAG and tool-use model with open research weights", "source": "https://cohere.com/blog/command-r-plus-microsoft-azure" }, { "name": "Kandinsky 3.1", "org": "Sber", "country": "Russia", "date": "2024-04-04", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Sber's upgraded open text-to-image model with Kandinsky Flash distilled variant", "source": "https://huggingface.co/ai-forever/Kandinsky3.1" }, { "name": "Sailor", "org": "Sea AI Lab", "country": "Singapore", "date": "2024-04-04", "precision": "day", "category": "language", "open_weights": true, "params": "0.5B-7B", "note": "Open LLMs tailored for Southeast Asian languages", "source": "https://arxiv.org/abs/2404.03608" }, { "name": "Open-Sora-Plan v1.0", "org": "Peking University", "country": "China", "date": "2024-04-07", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "PKU-Yuan Group's open Sora reproduction effort", "source": "https://github.com/PKU-YuanGroup/Open-Sora-Plan" }, { "name": "Ferret-UI", "org": "Apple", "country": "USA", "date": "2024-04-08", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Multimodal LLM grounded in mobile UI screen understanding", "source": "https://arxiv.org/abs/2404.05719" }, { "name": "Stable LM 2 12B", "org": "Stability AI", "country": "UK", "date": "2024-04-08", "precision": "day", "category": "language", "open_weights": true, "params": "12B", "note": "Multilingual 12B model across seven European languages", "source": "https://stability.ai/news-updates/introducing-stable-lm-2-12b" }, { "name": "YaART", "org": "Yandex", "country": "Russia", "date": "2024-04-08", "precision": "day", "category": "image", "open_weights": false, "params": "2.3B", "note": "Yandex's cascaded diffusion text-to-image model tuned with RLHF", "source": "https://arxiv.org/abs/2404.05666" }, { "name": "CodeGemma", "org": "Google DeepMind", "country": "UK", "date": "2024-04-09", "precision": "day", "category": "code", "open_weights": true, "params": "2B, 7B", "note": "Gemma-based open models for code completion and generation", "source": "https://developers.googleblog.com/en/gemma-family-expands-with-models-tailored-for-developers-and-researchers/" }, { "name": "GPT-4 Turbo with Vision", "org": "OpenAI", "country": "USA", "date": "2024-04-09", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Generally available GPT-4 Turbo update with vision built in", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "RecurrentGemma", "org": "Google DeepMind", "country": "UK", "date": "2024-04-09", "precision": "day", "category": "language", "open_weights": true, "params": "2B", "note": "Open Griffin-architecture model with efficient long-sequence inference", "source": "https://developers.googleblog.com/en/gemma-family-expands-with-models-tailored-for-developers-and-researchers/" }, { "name": "InstantMesh", "org": "Tencent", "country": "China", "date": "2024-04-10", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Single image to 3D mesh in seconds via sparse-view reconstruction", "source": "https://arxiv.org/abs/2404.07191" }, { "name": "Mixtral 8x22B", "org": "Mistral AI", "country": "France", "date": "2024-04-10", "precision": "day", "category": "language", "open_weights": true, "params": "141B (39B active)", "note": "Open sparse MoE first released via a BitTorrent link", "source": "https://en.wikipedia.org/wiki/Mistral_AI" }, { "name": "Udio", "org": "Udio", "country": "USA", "date": "2024-04-10", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Text-to-music generator from ex-DeepMind researchers, public beta", "source": "https://en.wikipedia.org/wiki/Udio" }, { "name": "Zephyr 141B-A39B", "org": "Hugging Face", "country": "USA", "date": "2024-04-10", "precision": "day", "category": "language", "open_weights": true, "params": "141B (39B active)", "note": "ORPO-aligned fine-tune of Mixtral 8x22B", "source": "https://huggingface.co/HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1" }, { "name": "Rerank 3", "org": "Cohere", "country": "Canada", "date": "2024-04-11", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Enterprise search reranking model", "source": "https://cohere.com/blog/rerank-3" }, { "name": "Grok-1.5V", "org": "xAI", "country": "USA", "date": "2024-04-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "xAI's first multimodal Grok, understanding images and documents", "source": "https://x.ai/news/grok-1.5v" }, { "name": "MiniCPM-V 2.0", "org": "ModelBest", "country": "China", "date": "2024-04-12", "precision": "day", "category": "language", "open_weights": true, "params": "2.8B", "note": "Edge vision-language model rivaling Gemini Pro on scene text", "source": "https://github.com/OpenBMB/MiniCPM-V" }, { "name": "Idefics2", "org": "Hugging Face", "country": "USA", "date": "2024-04-15", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Open 8B vision-language model with stronger OCR", "source": "https://huggingface.co/blog/idefics2" }, { "name": "Reka Core", "org": "Reka AI", "country": "USA", "date": "2024-04-15", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Reka's frontier multimodal model accepting image, video and audio", "source": "https://www.reka.ai/news/reka-core-our-frontier-class-multimodal-language-model" }, { "name": "WizardLM-2", "org": "Microsoft", "country": "USA", "date": "2024-04-15", "precision": "day", "category": "language", "open_weights": true, "params": "7B-8x22B", "note": "Released then briefly pulled for missing toxicity testing", "source": "https://wizardlm.github.io/WizardLM2/" }, { "name": "CodeQwen1.5", "org": "Alibaba", "country": "China", "date": "2024-04-16", "precision": "day", "category": "code", "open_weights": true, "params": "7B", "note": "Code-specialized Qwen model with 64K context", "source": "https://qwenlm.github.io/blog/codeqwen1.5/" }, { "name": "Snowflake Arctic Embed", "org": "Snowflake", "country": "USA", "date": "2024-04-16", "precision": "day", "category": "language", "open_weights": true, "params": "22M-335M", "note": "Open text-embedding family optimized for retrieval", "source": "https://www.snowflake.com/en/blog/introducing-snowflake-arctic-embed-snowflakes-state-of-the-art-text-embedding-family-of-models/" }, { "name": "VASA-1", "org": "Microsoft", "country": "USA", "date": "2024-04-16", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Lifelike real-time talking faces from one photo and audio", "source": "https://www.microsoft.com/en-us/research/project/vasa-1/" }, { "name": "Zamba", "org": "Zyphra", "country": "USA", "date": "2024-04-16", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Mamba plus shared-attention hybrid 7B model", "source": "https://www.zyphra.com/post/zamba" }, { "name": "abab6.5", "org": "MiniMax", "country": "China", "date": "2024-04-17", "precision": "day", "category": "language", "open_weights": false, "params": "~1T (MoE)", "note": "MiniMax's trillion-parameter-class MoE LLM series", "source": "https://www.minimax.io/news/abab65-series" }, { "name": "LINGO-2", "org": "Wayve", "country": "UK", "date": "2024-04-17", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Vision-language-action driving model narrating its actions on public roads", "source": "https://wayve.ai/thinking/lingo-2-driving-with-language/" }, { "name": "OLMo 1.7-7B", "org": "Allen Institute for AI", "country": "USA", "date": "2024-04-17", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "24-point MMLU gain over the original fully open OLMo", "source": "https://allenai.org/blog/olmo-1-7-7b-a-24-point-improvement-on-mmlu-92b43f7d269d" }, { "name": "Tiangong 3.0", "org": "Kunlun Tech", "country": "China", "date": "2024-04-17", "precision": "day", "category": "language", "open_weights": false, "params": "400B (MoE)", "note": "Kunlun's 400B MoE flagship opened to public beta", "source": "https://www.leiphone.com/category/industrynews/aKm3DiETl4rw1A6G.html" }, { "name": "Tiangong SkyMusic", "org": "Kunlun Tech", "country": "China", "date": "2024-04-17", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Kunlun's text-to-music model, launched in public beta", "source": "https://www.leiphone.com/category/industrynews/aKm3DiETl4rw1A6G.html" }, { "name": "Llama 3", "org": "Meta", "country": "USA", "date": "2024-04-18", "precision": "day", "category": "language", "open_weights": true, "params": "8B, 70B", "note": "Trained on 15T tokens; powered the relaunched Meta AI assistant", "source": "https://ai.meta.com/blog/meta-llama-3/" }, { "name": "Hyper-SD", "org": "ByteDance", "country": "China", "date": "2024-04-21", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Trajectory-segmented consistency distillation for few-step diffusion", "source": "https://arxiv.org/abs/2404.13686" }, { "name": "OpenELM", "org": "Apple", "country": "USA", "date": "2024-04-22", "precision": "day", "category": "language", "open_weights": true, "params": "270M-3B", "note": "Apple's open efficient LMs released with full training framework", "source": "https://arxiv.org/abs/2404.14619" }, { "name": "Firefly Image 3", "org": "Adobe", "country": "USA", "date": "2024-04-23", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Adobe's third-generation commercially safe image foundation model", "source": "https://news.adobe.com/news/news-details/2024/adobe-introduces-firefly-image-3-foundation-model-to-take-creative-exploration-and-ideation-to-new-heights" }, { "name": "Phi-3-medium", "org": "Microsoft", "country": "USA", "date": "2024-04-23", "precision": "day", "category": "language", "open_weights": true, "params": "14B", "note": "Announced April 23; weights released May 21", "source": "https://azure.microsoft.com/en-us/blog/introducing-phi-3-redefining-whats-possible-with-slms/" }, { "name": "Phi-3-mini", "org": "Microsoft", "country": "USA", "date": "2024-04-23", "precision": "day", "category": "language", "open_weights": true, "params": "3.8B", "note": "Phone-runnable small model rivaling GPT-3.5 and Mixtral", "source": "https://azure.microsoft.com/en-us/blog/introducing-phi-3-redefining-whats-possible-with-slms/" }, { "name": "Phi-3-small", "org": "Microsoft", "country": "USA", "date": "2024-04-23", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Announced April 23; weights released May 21", "source": "https://azure.microsoft.com/en-us/blog/introducing-phi-3-redefining-whats-possible-with-slms/" }, { "name": "SenseNova 5.0", "org": "SenseTime", "country": "China", "date": "2024-04-23", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "SenseTime flagship LLM; launch sent shares up over 30%", "source": "https://www.scmp.com/tech/tech-trends/article/3260203/chinese-ai-giant-sensetime-suspends-trading-shares-surge-more-30-after-launch-updated-large-language" }, { "name": "Snowflake Arctic", "org": "Snowflake", "country": "USA", "date": "2024-04-24", "precision": "day", "category": "language", "open_weights": true, "params": "480B (17B active)", "note": "Dense-MoE hybrid enterprise LLM under Apache 2.0", "source": "https://www.snowflake.com/en/blog/arctic-open-efficient-foundation-language-models-snowflake/" }, { "name": "InternVL 1.5", "org": "Shanghai AI Lab", "country": "China", "date": "2024-04-25", "precision": "day", "category": "language", "open_weights": true, "params": "26B", "note": "Open multimodal suite closing the gap to GPT-4V", "source": "https://arxiv.org/abs/2404.16821" }, { "name": "Qwen1.5-110B", "org": "Alibaba", "country": "China", "date": "2024-04-25", "precision": "day", "category": "language", "open_weights": true, "params": "110B", "note": "First 100B+ open model in the Qwen1.5 series", "source": "https://qwenlm.github.io/blog/qwen1.5-110b/" }, { "name": "Vidu", "org": "ShengShu", "country": "China", "date": "2024-04-27", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "China's first Sora-class long-video model, built with Tsinghua", "source": "https://interestingengineering.com/innovation/china-openai-sora-rival-vidu" }, { "name": "Med-Gemini", "org": "Google", "country": "USA", "date": "2024-04-29", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Medicine-specialized Gemini family with state-of-the-art MedQA", "source": "https://arxiv.org/abs/2404.18416" }, { "name": "ChatTTS", "org": "2noise", "country": "China", "date": "2024-05-01", "precision": "month", "category": "audio", "open_weights": true, "params": "", "note": "Viral open TTS model optimized for conversational dialogue", "source": "https://github.com/2noise/ChatTTS" }, { "name": "LLM360 K2", "org": "MBZUAI", "country": "UAE", "date": "2024-05-01", "precision": "month", "category": "language", "open_weights": true, "params": "65B", "note": "Fully transparent 65B model with open data and checkpoints", "source": "https://huggingface.co/LLM360/K2" }, { "name": "Qwen-Max-0428", "org": "Alibaba", "country": "China", "date": "2024-05-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Proprietary Qwen flagship that reached Chatbot Arena top 10", "source": "https://qwenlm.github.io/blog/qwen-max-0428/" }, { "name": "VILA 1.5", "org": "NVIDIA", "country": "USA", "date": "2024-05-01", "precision": "month", "category": "language", "open_weights": true, "params": "3B-40B", "note": "Efficient visual language models with video understanding, deployable to edge", "source": "https://github.com/NVlabs/VILA" }, { "name": "Jamba-Instruct", "org": "AI21 Labs", "country": "Israel", "date": "2024-05-02", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Instruction-tuned Jamba with 256K context for enterprise", "source": "https://www.ai21.com/blog/announcing-jamba-instruct" }, { "name": "StoryDiffusion", "org": "ByteDance", "country": "China", "date": "2024-05-02", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Consistent self-attention for long-range image and video storytelling", "source": "https://arxiv.org/abs/2405.01434" }, { "name": "DeepSeek-V2", "org": "DeepSeek", "country": "China", "date": "2024-05-06", "precision": "day", "category": "language", "open_weights": true, "params": "236B (21B active)", "note": "Introduced MLA attention; low pricing sparked China's LLM price war", "source": "https://github.com/deepseek-ai/DeepSeek-V2" }, { "name": "Granite Code", "org": "IBM", "country": "USA", "date": "2024-05-06", "precision": "day", "category": "code", "open_weights": true, "params": "3B-34B", "note": "Apache-2.0 code models trained on 116 programming languages", "source": "https://research.ibm.com/blog/granite-code-models-open-source" }, { "name": "Amazon Titan Text Premier", "org": "Amazon", "country": "USA", "date": "2024-05-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Most capable Titan LLM, built for RAG and agents on Bedrock", "source": "https://aws.amazon.com/blogs/aws/build-rag-and-agent-based-generative-ai-applications-with-new-amazon-titan-text-premier-model-available-in-amazon-bedrock/" }, { "name": "xLSTM", "org": "NXAI / JKU Linz", "country": "Austria", "date": "2024-05-07", "precision": "day", "category": "language", "open_weights": false, "params": "1.3B", "note": "Hochreiter revives LSTM with exponential gating to rival Transformers", "source": "https://arxiv.org/abs/2405.04517" }, { "name": "AlphaFold 3", "org": "Google DeepMind", "country": "UK", "date": "2024-05-08", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Predicts structures and interactions of proteins, DNA, RNA, ligands", "source": "https://blog.google/technology/ai/google-deepmind-isomorphic-alphafold-3-ai-model/" }, { "name": "MatterSim", "org": "Microsoft", "country": "USA", "date": "2024-05-08", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Atomistic materials model across elements, temperatures and pressures", "source": "https://arxiv.org/abs/2405.04967" }, { "name": "Fugaku-LLM", "org": "Fujitsu / RIKEN", "country": "Japan", "date": "2024-05-10", "precision": "day", "category": "language", "open_weights": true, "params": "13B", "note": "Japanese LLM trained on the CPU-based Fugaku supercomputer", "source": "https://www.fujitsu.com/global/about/resources/news/press-releases/2024/0510-01.html" }, { "name": "Falcon 2 11B", "org": "TII", "country": "UAE", "date": "2024-05-13", "precision": "day", "category": "language", "open_weights": true, "params": "11B", "note": "TII's second-generation open model trained on 5.5T tokens", "source": "https://falconllm.tii.ae/falcon-2.html" }, { "name": "Falcon 2 11B VLM", "org": "TII", "country": "UAE", "date": "2024-05-13", "precision": "day", "category": "language", "open_weights": true, "params": "11B", "note": "TII's first vision-language Falcon", "source": "https://huggingface.co/blog/falcon2-11b" }, { "name": "GPT-4o", "org": "OpenAI", "country": "USA", "date": "2024-05-13", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Natively omni model reasoning across text, audio and vision in real time", "source": "https://openai.com/index/hello-gpt-4o/" }, { "name": "Yi-1.5", "org": "01.AI", "country": "China", "date": "2024-05-13", "precision": "day", "category": "language", "open_weights": true, "params": "6B, 9B, 34B", "note": "Upgraded open Yi with stronger coding, math and reasoning", "source": "https://github.com/01-ai/Yi-1.5" }, { "name": "Yi-Large", "org": "01.AI", "country": "China", "date": "2024-05-13", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "01.AI's proprietary flagship LLM", "source": "https://www.trendingtopics.eu/01-ai-chinese-ai-startup-shakes-up-the-world-of-llms/" }, { "name": "Gemini 1.5 Flash", "org": "Google DeepMind", "country": "UK", "date": "2024-05-14", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Fast, lightweight 1M-context model distilled from 1.5 Pro", "source": "https://blog.google/technology/ai/google-gemini-update-flash-ai-assistant-io-2024/" }, { "name": "Gemma 2", "org": "Google DeepMind", "country": "UK", "date": "2024-05-14", "precision": "day", "category": "language", "open_weights": true, "params": "9B, 27B", "note": "Announced at I/O; 9B and 27B weights released June 27", "source": "https://developers.googleblog.com/en/gemma-family-and-toolkit-expansion-io-2024/" }, { "name": "Hunyuan-DiT", "org": "Tencent", "country": "China", "date": "2024-05-14", "precision": "day", "category": "image", "open_weights": true, "params": "1.5B", "note": "Bilingual diffusion transformer with fine-grained Chinese understanding", "source": "https://arxiv.org/abs/2405.08748" }, { "name": "Imagen 3", "org": "Google DeepMind", "country": "UK", "date": "2024-05-14", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Google's highest-quality text-to-image model, unveiled at I/O", "source": "https://blog.google/technology/ai/google-generative-ai-veo-imagen-3/" }, { "name": "LearnLM", "org": "Google", "country": "USA", "date": "2024-05-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Gemini-based models fine-tuned for learning and tutoring", "source": "https://storage.googleapis.com/deepmind-media/LearnLM/LearnLM_paper.pdf" }, { "name": "PaliGemma", "org": "Google DeepMind", "country": "UK", "date": "2024-05-14", "precision": "day", "category": "language", "open_weights": true, "params": "3B", "note": "Open vision-language model built for fine-tuning on vision tasks", "source": "https://developers.googleblog.com/en/gemma-family-and-toolkit-expansion-io-2024/" }, { "name": "Project Astra", "org": "Google DeepMind", "country": "UK", "date": "2024-05-14", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Prototype universal assistant with real-time video understanding", "source": "https://blog.google/technology/ai/google-gemini-update-flash-ai-assistant-io-2024/" }, { "name": "Veo", "org": "Google DeepMind", "country": "UK", "date": "2024-05-14", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "1080p text-to-video model generating clips beyond a minute", "source": "https://blog.google/technology/ai/google-generative-ai-veo-imagen-3/" }, { "name": "Doubao", "org": "ByteDance", "country": "China", "date": "2024-05-15", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Doubao Pro/Lite family at rock-bottom prices; fueled China's LLM price war", "source": "https://www.scmp.com/tech/big-tech/article/3262781/tiktok-owner-bytedance-launches-low-cost-doubao-ai-models-enterprises-initiating-price-war-crowded" }, { "name": "Chameleon", "org": "Meta", "country": "USA", "date": "2024-05-16", "precision": "day", "category": "multimodal", "open_weights": true, "params": "7B, 34B", "note": "Early-fusion token-based mixed-modal text and image model", "source": "https://arxiv.org/abs/2405.09818" }, { "name": "DeepSeek-V2-Lite", "org": "DeepSeek", "country": "China", "date": "2024-05-16", "precision": "day", "category": "language", "open_weights": true, "params": "15.7B (2.4B active)", "note": "Small MLA/MoE variant deployable on a single 40GB GPU", "source": "https://github.com/deepseek-ai/DeepSeek-V2" }, { "name": "Grounding DINO 1.5", "org": "IDEA Research", "country": "China", "date": "2024-05-16", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Open-set object detection with Pro and Edge variants", "source": "https://arxiv.org/abs/2405.10300" }, { "name": "Aurora", "org": "Microsoft", "country": "USA", "date": "2024-05-20", "precision": "day", "category": "science", "open_weights": true, "params": "1.3B", "note": "Atmosphere foundation model for weather and air-pollution forecasting", "source": "https://arxiv.org/abs/2405.13063" }, { "name": "CogVLM2", "org": "Zhipu AI", "country": "China", "date": "2024-05-20", "precision": "day", "category": "language", "open_weights": true, "params": "19B", "note": "Open Llama-3-based vision-language model from Zhipu and Tsinghua", "source": "https://github.com/THUDM/CogVLM2" }, { "name": "DIAMOND", "org": "University of Geneva", "country": "Switzerland", "date": "2024-05-20", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Diffusion world model used to train RL agents on Atari", "source": "https://arxiv.org/abs/2405.12399" }, { "name": "MiniCPM-Llama3-V 2.5", "org": "ModelBest", "country": "China", "date": "2024-05-20", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Phone-deployable MLLM claiming GPT-4V-level performance", "source": "https://github.com/OpenBMB/MiniCPM-V" }, { "name": "BiomedParse", "org": "Microsoft", "country": "USA", "date": "2024-05-21", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Biomedical foundation model for joint segmentation, detection and recognition", "source": "https://arxiv.org/abs/2405.12971" }, { "name": "Phi Silica", "org": "Microsoft", "country": "USA", "date": "2024-05-21", "precision": "day", "category": "language", "open_weights": false, "params": "3.3B", "note": "On-device small language model built for Copilot+ PC NPUs", "source": "https://blogs.windows.com/windowsdeveloper/2024/05/21/unlock-a-new-era-of-innovation-with-windows-copilot-runtime-and-copilot-pcs/" }, { "name": "Phi-3-vision", "org": "Microsoft", "country": "USA", "date": "2024-05-21", "precision": "day", "category": "language", "open_weights": true, "params": "4.2B", "note": "First multimodal Phi model, reading charts and images", "source": "https://azure.microsoft.com/en-us/blog/new-models-added-to-the-phi-3-family-available-on-microsoft-azure/" }, { "name": "Baichuan 4", "org": "Baichuan", "country": "China", "date": "2024-05-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Topped SuperCLUE Chinese benchmark; launched with Baichuan's first AI assistant", "source": "https://zhidx.com/p/426526.html" }, { "name": "Mistral 7B v0.3", "org": "Mistral AI", "country": "France", "date": "2024-05-22", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Extended vocabulary and native function calling", "source": "https://huggingface.co/mistralai/Mistral-7B-v0.3" }, { "name": "Prov-GigaPath", "org": "Microsoft", "country": "USA", "date": "2024-05-22", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Whole-slide digital pathology foundation model, with Providence", "source": "https://www.microsoft.com/en-us/research/blog/gigapath-whole-slide-foundation-model-for-digital-pathology/" }, { "name": "Aya 23", "org": "Cohere", "country": "Canada", "date": "2024-05-23", "precision": "day", "category": "language", "open_weights": true, "params": "8B, 35B", "note": "Cohere For AI open multilingual models covering 23 languages", "source": "https://arxiv.org/abs/2405.15032" }, { "name": "DeepSeek-Prover", "org": "DeepSeek", "country": "China", "date": "2024-05-23", "precision": "day", "category": "science", "open_weights": true, "params": "7B", "note": "Lean 4 theorem prover trained on large-scale synthetic data", "source": "https://arxiv.org/abs/2405.14333" }, { "name": "Golden Gate Claude", "org": "Anthropic", "country": "USA", "date": "2024-05-23", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "24-hour interpretability demo: Claude with Golden Gate Bridge feature amplified", "source": "https://www.anthropic.com/news/golden-gate-claude" }, { "name": "YOLOv10", "org": "Tsinghua University", "country": "China", "date": "2024-05-23", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "NMS-free end-to-end real-time object detection", "source": "https://arxiv.org/abs/2405.14458" }, { "name": "NV-Embed", "org": "NVIDIA", "country": "USA", "date": "2024-05-27", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Mistral-based generalist embedding model that topped MTEB", "source": "https://arxiv.org/abs/2405.17428" }, { "name": "YandexGPT 3 Lite", "org": "Yandex", "country": "Russia", "date": "2024-05-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Lighter, faster tier of YandexGPT 3", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Codestral", "org": "Mistral AI", "country": "France", "date": "2024-05-29", "precision": "day", "category": "code", "open_weights": true, "params": "22B", "note": "Mistral's first code model, fluent in 80+ programming languages", "source": "https://mistral.ai/news/codestral/" }, { "name": "ElevenLabs Text to Sound Effects", "org": "ElevenLabs", "country": "USA", "date": "2024-05-31", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Generates sound effects from text prompts", "source": "https://elevenlabs.io/blog/sound-effects-are-here" }, { "name": "Mamba-2", "org": "Princeton University / Carnegie Mellon University", "country": "USA", "date": "2024-05-31", "precision": "day", "category": "language", "open_weights": true, "params": "up to 2.7B", "note": "Structured state-space duality unifying SSMs and attention", "source": "https://arxiv.org/abs/2405.21060" }, { "name": "Sonic", "org": "Cartesia", "country": "USA", "date": "2024-05-31", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Low-latency state-space voice model for lifelike speech", "source": "https://www.cartesia.ai/blog/sonic" }, { "name": "Leonardo Phoenix", "org": "Leonardo AI", "country": "Australia", "date": "2024-06-01", "precision": "month", "category": "image", "open_weights": false, "params": "", "note": "Leonardo's first in-house foundation image model", "source": "https://en.wikipedia.org/wiki/Leonardo.ai" }, { "name": "Skywork-MoE", "org": "Kunlun Tech", "country": "China", "date": "2024-06-03", "precision": "day", "category": "language", "open_weights": true, "params": "146B (22B active)", "note": "Open 146B mixture-of-experts model from Kunlun's Skywork team", "source": "https://github.com/SkyworkAI/Skywork-MoE" }, { "name": "Seed-TTS", "org": "ByteDance", "country": "China", "date": "2024-06-04", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech generation family approaching human-level naturalness", "source": "https://arxiv.org/abs/2406.02430" }, { "name": "GLM-4-9B", "org": "Zhipu AI", "country": "China", "date": "2024-06-05", "precision": "day", "category": "language", "open_weights": true, "params": "9B", "note": "Open GLM-4 series incl. 1M-context chat and GLM-4V-9B vision", "source": "https://github.com/THUDM/GLM-4" }, { "name": "Nomic Embed Vision", "org": "Nomic AI", "country": "USA", "date": "2024-06-05", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Image embeddings sharing Nomic Embed's text latent space", "source": "https://huggingface.co/nomic-ai/nomic-embed-vision-v1.5" }, { "name": "Stable Audio Open", "org": "Stability AI", "country": "UK", "date": "2024-06-05", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open-weights model for short samples and sound effects", "source": "https://stability.ai/news-updates/introducing-stable-audio-open" }, { "name": "Kling", "org": "Kuaishou", "country": "China", "date": "2024-06-06", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "First Chinese Sora-class video model open to public testing", "source": "https://www.maginative.com/article/kuaishou-unveils-kling-a-text-to-video-model-to-challenge-openais-sora/" }, { "name": "Qwen2", "org": "Alibaba", "country": "China", "date": "2024-06-07", "precision": "day", "category": "language", "open_weights": true, "params": "0.5B-72B", "note": "Open family up to 72B, including a 57B-A14B MoE", "source": "https://qwenlm.github.io/blog/qwen2/" }, { "name": "VALL-E 2", "org": "Microsoft", "country": "USA", "date": "2024-06-08", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Codec language model claiming human parity in zero-shot TTS", "source": "https://arxiv.org/abs/2406.05370" }, { "name": "Apple Foundation Models", "org": "Apple", "country": "USA", "date": "2024-06-10", "precision": "day", "category": "language", "open_weights": false, "params": "~3B on-device + server", "note": "On-device and server models powering Apple Intelligence, unveiled at WWDC", "source": "https://machinelearning.apple.com/research/introducing-apple-foundation-models" }, { "name": "Dream Machine", "org": "Luma AI", "country": "USA", "date": "2024-06-12", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Luma's text-to-video model, freely available at launch", "source": "https://en.wikipedia.org/wiki/Dream_Machine_(text-to-video_model)" }, { "name": "Shutterstock ImageAI", "org": "Databricks", "country": "USA", "date": "2024-06-12", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Image model trained only on Shutterstock's licensed library", "source": "https://www.databricks.com/company/newsroom/press-releases/introducing-shutterstock-imageai-powered-databricks-image" }, { "name": "Stable Diffusion 3 Medium", "org": "Stability AI", "country": "UK", "date": "2024-06-12", "precision": "day", "category": "image", "open_weights": true, "params": "2B", "note": "First open-weights Stable Diffusion 3 model", "source": "https://stability.ai/news-updates/stable-diffusion-3-medium" }, { "name": "Depth Anything V2", "org": "ByteDance", "country": "China", "date": "2024-06-13", "precision": "day", "category": "vision", "open_weights": true, "params": "25M-1.3B", "note": "Finer, more robust depth estimation via synthetic-data training", "source": "https://arxiv.org/abs/2406.09414" }, { "name": "Hallo", "org": "Fudan University", "country": "China", "date": "2024-06-13", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Hierarchical audio-driven portrait image animation", "source": "https://arxiv.org/abs/2406.08801" }, { "name": "OpenVLA", "org": "Stanford University", "country": "USA", "date": "2024-06-13", "precision": "day", "category": "robotics", "open_weights": true, "params": "7B", "note": "Open vision-language-action model trained on 970K robot episodes", "source": "https://arxiv.org/abs/2406.09246" }, { "name": "Nemotron-4 340B", "org": "NVIDIA", "country": "USA", "date": "2024-06-14", "precision": "day", "category": "language", "open_weights": true, "params": "340B", "note": "Open base, instruct and reward models for synthetic data generation", "source": "https://blogs.nvidia.com/blog/nemotron-4-synthetic-data-generation-llm-training/" }, { "name": "PLaMo-100B", "org": "Preferred Networks", "country": "Japan", "date": "2024-06-14", "precision": "day", "category": "language", "open_weights": false, "params": "100B", "note": "Japanese 100B-parameter LLM pretrained from scratch", "source": "https://preferred.jp/ja/blog/tech/plamo-100b/" }, { "name": "JASCO", "org": "Meta", "country": "USA", "date": "2024-06-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Text-to-music conditioned on chords, melody and drums", "source": "https://arxiv.org/abs/2406.10970" }, { "name": "DeepSeek-Coder-V2", "org": "DeepSeek", "country": "China", "date": "2024-06-17", "precision": "day", "category": "code", "open_weights": true, "params": "236B (21B active)", "note": "Open MoE code model rivaling GPT-4 Turbo on coding benchmarks", "source": "https://arxiv.org/abs/2406.11931" }, { "name": "Gen-3 Alpha", "org": "Runway", "country": "USA", "date": "2024-06-17", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Higher-fidelity video model on Runway's new training infrastructure", "source": "https://runwayml.com/research/introducing-gen-3-alpha" }, { "name": "PRISM-1", "org": "Wayve", "country": "UK", "date": "2024-06-17", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "4D reconstruction of dynamic driving scenes for simulation", "source": "https://wayve.ai/thinking/prism-1/" }, { "name": "V2A", "org": "Google DeepMind", "country": "UK", "date": "2024-06-17", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Video-to-audio model generating soundtracks and dialogue for silent video", "source": "https://deepmind.google/discover/blog/generating-audio-for-video/" }, { "name": "Code Droid", "org": "Factory", "country": "USA", "date": "2024-06-18", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Autonomous coding agent with strong SWE-bench results", "source": "https://factory.com/news/code-droid-technical-report" }, { "name": "Claude 3.5 Sonnet", "org": "Anthropic", "country": "USA", "date": "2024-06-20", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Beat Claude 3 Opus at mid-tier speed; debuted Artifacts", "source": "https://www.anthropic.com/news/claude-3-5-sonnet" }, { "name": "Pangu 5.0", "org": "Huawei", "country": "China", "date": "2024-06-21", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Huawei's industry-focused model family unveiled at HDC 2024", "source": "https://www.huaweicentral.com/huawei-cloud-unveils-pangu-large-model-5-0/" }, { "name": "Cambrian-1", "org": "New York University", "country": "USA", "date": "2024-06-24", "precision": "day", "category": "language", "open_weights": true, "params": "8B-34B", "note": "Fully open, vision-centric multimodal LLM study", "source": "https://arxiv.org/abs/2406.16860" }, { "name": "ESM3", "org": "EvolutionaryScale", "country": "USA", "date": "2024-06-25", "precision": "day", "category": "science", "open_weights": true, "params": "98B (1.4B open)", "note": "Protein language model that designed novel fluorescent protein esmGFP", "source": "https://www.evolutionaryscale.ai/blog/esm3-release" }, { "name": "E2 TTS", "org": "Microsoft", "country": "USA", "date": "2024-06-26", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Fully non-autoregressive zero-shot TTS", "source": "https://arxiv.org/abs/2406.18009" }, { "name": "CriticGPT", "org": "OpenAI", "country": "USA", "date": "2024-06-27", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "GPT-4-based critic catching ChatGPT code errors for RLHF", "source": "https://arxiv.org/abs/2407.00215" }, { "name": "LLM Compiler", "org": "Meta", "country": "USA", "date": "2024-06-27", "precision": "day", "category": "code", "open_weights": true, "params": "7B, 13B", "note": "Code Llama-based models for compiler IR and optimization", "source": "https://arxiv.org/abs/2407.02524" }, { "name": "Spark 4.0", "org": "iFlytek", "country": "China", "date": "2024-06-27", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Trained on domestic compute cluster; claimed to beat GPT-4 Turbo in areas", "source": "https://news.mydrivers.com/1/988/988242.htm" }, { "name": "ERNIE 4.0 Turbo", "org": "Baidu", "country": "China", "date": "2024-06-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Faster ERNIE 4.0 variant; Ernie Bot users hit 300M", "source": "https://www.reuters.com/technology/artificial-intelligence/baidu-launches-upgraded-ai-model-says-user-base-hits-300-mln-2024-06-28/" }, { "name": "CogVideoX", "org": "Zhipu AI", "country": "China", "date": "2024-07-01", "precision": "month", "category": "video", "open_weights": true, "params": "2B", "note": "Powers Zhipu's Ying video generator (July); 2B weights opened August 6", "source": "https://github.com/zai-org/CogVideo" }, { "name": "Meta 3D Gen", "org": "Meta", "country": "USA", "date": "2024-07-02", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Text-to-3D pipeline producing textured meshes in under a minute", "source": "https://arxiv.org/abs/2407.02599" }, { "name": "InternLM-XComposer-2.5", "org": "Shanghai AI Lab", "country": "China", "date": "2024-07-03", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Open long-context vision-language model for images, video and composition", "source": "https://github.com/InternLM/InternLM-XComposer" }, { "name": "InternLM2.5", "org": "Shanghai AI Lab", "country": "China", "date": "2024-07-03", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Open 7B with 1M-token chat variant; 1.8B and 20B followed August 1", "source": "https://github.com/InternLM/InternLM" }, { "name": "LivePortrait", "org": "Kuaishou", "country": "China", "date": "2024-07-03", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Popular open portrait animation model from the Kling team", "source": "https://github.com/KlingAIResearch/LivePortrait" }, { "name": "Moshi", "org": "Kyutai", "country": "France", "date": "2024-07-03", "precision": "day", "category": "audio", "open_weights": true, "params": "7B", "note": "First real-time full-duplex spoken dialogue model; weights opened September 2024", "source": "https://kyutai.org/blog/2024-07-03-meet-moshi/" }, { "name": "Tele-FLM-1T", "org": "BAAI", "country": "China", "date": "2024-07-03", "precision": "day", "category": "language", "open_weights": true, "params": "1T", "note": "Open trillion-parameter dense checkpoint from BAAI and China Telecom's TeleAI", "source": "https://arxiv.org/abs/2407.02783" }, { "name": "CosyVoice", "org": "Alibaba", "country": "China", "date": "2024-07-04", "precision": "day", "category": "audio", "open_weights": true, "params": "300M", "note": "Open multilingual zero-shot TTS released in the FunAudioLLM suite", "source": "https://arxiv.org/abs/2407.04051" }, { "name": "InternVL2", "org": "Shanghai AI Lab", "country": "China", "date": "2024-07-04", "precision": "day", "category": "language", "open_weights": true, "params": "1B-76B", "note": "Open vision-language model family spanning 1B to 76B parameters", "source": "https://internvl.github.io/blog/2024-07-02-InternVL-2.0/" }, { "name": "SenseVoice", "org": "Alibaba", "country": "China", "date": "2024-07-04", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open multilingual speech recognition, emotion and audio-event model from FunAudioLLM", "source": "https://arxiv.org/abs/2407.04051" }, { "name": "Step-1.5V", "org": "StepFun", "country": "China", "date": "2024-07-04", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Multimodal understanding model launched at WAIC 2024", "source": "https://pandaily.com/stepfun-releases-three-large-models-of-the-step-series" }, { "name": "Step-1X", "org": "StepFun", "country": "China", "date": "2024-07-04", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Image generation model launched at WAIC 2024", "source": "https://pandaily.com/stepfun-releases-three-large-models-of-the-step-series" }, { "name": "CodeGeeX4-ALL-9B", "org": "Zhipu AI", "country": "China", "date": "2024-07-05", "precision": "day", "category": "code", "open_weights": true, "params": "9B", "note": "Open multilingual code model for completion, chat and repository Q&A", "source": "https://github.com/THUDM/CodeGeeX4" }, { "name": "SenseNova 5.5", "org": "SenseTime", "country": "China", "date": "2024-07-05", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "WAIC 2024 upgrade, launched alongside real-time multimodal SenseNova 5o", "source": "https://www.scmp.com/tech/big-tech/article/3269387/chinas-ai-competition-deepens-sensetime-alibaba-claim-progress-ai-show" }, { "name": "Kolors", "org": "Kuaishou", "country": "China", "date": "2024-07-06", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open bilingual Chinese-English text-to-image diffusion model", "source": "https://github.com/Kwai-Kolors/Kolors" }, { "name": "AuraFlow", "org": "fal", "country": "USA", "date": "2024-07-12", "precision": "day", "category": "image", "open_weights": true, "params": "6.8B", "note": "Largest fully open flow-based text-to-image model at release (v0.1)", "source": "https://blog.fal.ai/auraflow/" }, { "name": "Qwen2-Audio", "org": "Alibaba", "country": "China", "date": "2024-07-15", "precision": "day", "category": "audio", "open_weights": true, "params": "8.2B", "note": "Open audio-language model for voice chat and audio analysis", "source": "https://arxiv.org/abs/2407.10759" }, { "name": "Codestral Mamba", "org": "Mistral AI", "country": "France", "date": "2024-07-16", "precision": "day", "category": "code", "open_weights": true, "params": "7B", "note": "Mamba2 state-space code model under Apache 2.0", "source": "https://mistral.ai/news/codestral-mamba" }, { "name": "DeepL next-gen LLM", "org": "DeepL", "country": "Germany", "date": "2024-07-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Translation-specialised LLM trained on proprietary translation data", "source": "https://www.deepl.com/en/blog/next-gen-language-model" }, { "name": "Mathstral", "org": "Mistral AI", "country": "France", "date": "2024-07-16", "precision": "day", "category": "science", "open_weights": true, "params": "7B", "note": "Open 7B model specialised for math and STEM reasoning", "source": "https://mistral.ai/news/mathstral" }, { "name": "SmolLM", "org": "Hugging Face", "country": "USA", "date": "2024-07-16", "precision": "day", "category": "language", "open_weights": true, "params": "135M-1.7B", "note": "Tiny open models (135M/360M/1.7B) trained on curated SmolLM-Corpus", "source": "https://huggingface.co/blog/smollm" }, { "name": "GPT-4o mini", "org": "OpenAI", "country": "USA", "date": "2024-07-18", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Replaced GPT-3.5 Turbo; 82% MMLU at $0.15 per million input tokens", "source": "https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/" }, { "name": "Mistral NeMo", "org": "Mistral AI", "country": "France", "date": "2024-07-18", "precision": "day", "category": "language", "open_weights": true, "params": "12B", "note": "12B model built with NVIDIA; 128K context, new Tekken tokenizer", "source": "https://mistral.ai/news/mistral-nemo" }, { "name": "DCLM-7B", "org": "Apple", "country": "USA", "date": "2024-07-19", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Fully open 7B trained on curated DataComp-LM data; paper appeared June 17", "source": "https://venturebeat.com/ai/apple-shows-off-open-ai-prowess-new-models-outperform-mistral-and-hugging-face-offerings/" }, { "name": "Eleven Turbo v2.5", "org": "ElevenLabs", "country": "USA", "date": "2024-07-19", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Low-latency multilingual TTS, 3x faster across 32 languages", "source": "https://elevenlabs.io/blog/introducing-turbo-v25" }, { "name": "Minitron", "org": "NVIDIA", "country": "USA", "date": "2024-07-19", "precision": "day", "category": "language", "open_weights": true, "params": "4B, 8B", "note": "Pruned and distilled from Nemotron-4 15B", "source": "https://arxiv.org/abs/2407.14679" }, { "name": "Llama 3.1", "org": "Meta", "country": "USA", "date": "2024-07-23", "precision": "day", "category": "language", "open_weights": true, "params": "8B, 70B, 405B", "note": "405B was first open-weight model competitive with GPT-4-class frontier systems", "source": "https://huggingface.co/blog/llama31" }, { "name": "Udio v1.5", "org": "Udio", "country": "USA", "date": "2024-07-23", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music generator update with clearer audio and stem downloads", "source": "https://www.udio.com/blog/introducing-v1-5" }, { "name": "Mistral Large 2", "org": "Mistral AI", "country": "France", "date": "2024-07-24", "precision": "day", "category": "language", "open_weights": true, "params": "123B", "note": "123B flagship with 128K context and strong code, multilingual ability", "source": "https://mistral.ai/news/mistral-large-2407" }, { "name": "Stable Video 4D", "org": "Stability AI", "country": "UK", "date": "2024-07-24", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Generates novel-view videos of dynamic 3D objects from one video", "source": "https://stability.ai/news/stable-video-4d" }, { "name": "AlphaGeometry 2", "org": "Google DeepMind", "country": "UK", "date": "2024-07-25", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Gemini-trained geometry solver solving 83% of historical IMO geometry problems", "source": "https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/" }, { "name": "AlphaProof", "org": "Google DeepMind", "country": "UK", "date": "2024-07-25", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "RL plus Lean prover; with AlphaGeometry 2 reached IMO silver standard", "source": "https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/" }, { "name": "Zamba2-2.7B", "org": "Zyphra", "country": "USA", "date": "2024-07-28", "precision": "day", "category": "language", "open_weights": true, "params": "2.7B", "note": "Hybrid Mamba2-attention small model aimed at on-device inference", "source": "https://www.zyphra.com/post/zamba2-small" }, { "name": "SAM 2", "org": "Meta", "country": "USA", "date": "2024-07-29", "precision": "day", "category": "vision", "open_weights": true, "params": "224M", "note": "Promptable segmentation unified across images and video, Apache 2.0", "source": "https://ai.meta.com/research/publications/sam-2-segment-anything-in-images-and-videos/" }, { "name": "Llama-SEA-LION-v2-8B", "org": "AI Singapore", "country": "Singapore", "date": "2024-07-30", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Southeast Asian language model continued-pretrained from Llama 3", "source": "https://huggingface.co/aisingapore/Llama-SEA-LION-v2-8B" }, { "name": "Midjourney V6.1", "org": "Midjourney", "country": "USA", "date": "2024-07-30", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Improved coherence, textures and speed over V6", "source": "https://updates.midjourney.com/version-6-1/" }, { "name": "Gemma 2 2B", "org": "Google DeepMind", "country": "UK", "date": "2024-07-31", "precision": "day", "category": "language", "open_weights": true, "params": "2.6B", "note": "On-device Gemma 2 variant that beat GPT-3.5 on Chatbot Arena", "source": "https://developers.googleblog.com/en/smaller-safer-more-transparent-advancing-responsible-ai-with-gemma/" }, { "name": "ShieldGemma", "org": "Google DeepMind", "country": "UK", "date": "2024-07-31", "precision": "day", "category": "language", "open_weights": true, "params": "2B-27B", "note": "Gemma 2-based safety classifiers for filtering harmful inputs and outputs", "source": "https://huggingface.co/blog/gemma-july-update" }, { "name": "FLUX.1 [dev]", "org": "Black Forest Labs", "country": "Germany", "date": "2024-08-01", "precision": "day", "category": "image", "open_weights": true, "params": "12B", "note": "Guidance-distilled open-weight FLUX variant under non-commercial licence", "source": "https://huggingface.co/black-forest-labs/FLUX.1-dev" }, { "name": "FLUX.1 [pro]", "org": "Black Forest Labs", "country": "Germany", "date": "2024-08-01", "precision": "day", "category": "image", "open_weights": false, "params": "12B", "note": "Debut model of ex-Stable Diffusion researchers; new text-to-image state of the art", "source": "https://bfl.ai/blog/24-08-01-bfl" }, { "name": "FLUX.1 [schnell]", "org": "Black Forest Labs", "country": "Germany", "date": "2024-08-01", "precision": "day", "category": "image", "open_weights": true, "params": "12B", "note": "Few-step FLUX variant released under Apache 2.0", "source": "https://huggingface.co/black-forest-labs/FLUX.1-schnell" }, { "name": "Gemini 1.5 Pro Experimental 0801", "org": "Google DeepMind", "country": "UK", "date": "2024-08-01", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "First Gemini to reach #1 on LMSYS Chatbot Arena, passing GPT-4o", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "Idefics3", "org": "Hugging Face", "country": "USA", "date": "2024-08-01", "precision": "month", "category": "language", "open_weights": true, "params": "8B", "note": "Open Llama 3-based VLM trained only on public datasets", "source": "https://huggingface.co/HuggingFaceM4/Idefics3-8B-Llama3" }, { "name": "OmniParser", "org": "Microsoft", "country": "USA", "date": "2024-08-01", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Parses UI screenshots into structured elements for vision-only GUI agents", "source": "https://arxiv.org/abs/2408.00203" }, { "name": "Stable Fast 3D", "org": "Stability AI", "country": "UK", "date": "2024-08-01", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Single image to textured 3D asset in about half a second", "source": "https://stability.ai/news/introducing-stable-fast-3d" }, { "name": "Jais 70B", "org": "G42", "country": "UAE", "date": "2024-08-05", "precision": "day", "category": "language", "open_weights": true, "params": "70B", "note": "Flagship of 21 open Arabic-English Jais models from Inception/G42", "source": "https://www.g42.ai/resources/news/g42-launches-jais-70b-and-20-other-ai-models-champion-arabic-natural-language-processing" }, { "name": "LLaVA-OneVision", "org": "ByteDance", "country": "China", "date": "2024-08-06", "precision": "day", "category": "language", "open_weights": true, "params": "0.5B-72B", "note": "Open LMM transferring across single-image, multi-image and video tasks", "source": "https://arxiv.org/abs/2408.03326" }, { "name": "MiniCPM-V 2.6", "org": "OpenBMB", "country": "China", "date": "2024-08-06", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "8B on-device VLM claiming GPT-4V-level image and video understanding", "source": "https://huggingface.co/openbmb/MiniCPM-V-2_6" }, { "name": "Titan Image Generator v2", "org": "Amazon", "country": "USA", "date": "2024-08-06", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Adds image conditioning and background removal", "source": "https://aws.amazon.com/about-aws/whats-new/2024/08/titan-image-generator-v2-amazon-bedrock/" }, { "name": "Competitive robot table tennis agent", "org": "Google DeepMind", "country": "UK", "date": "2024-08-07", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "First learned robot agent reaching amateur human-level competitive table tennis", "source": "https://arxiv.org/abs/2408.03906" }, { "name": "EXAONE 3.0", "org": "LG AI Research", "country": "South Korea", "date": "2024-08-07", "precision": "day", "category": "language", "open_weights": true, "params": "7.8B", "note": "LG's first open-weight bilingual Korean-English LLM", "source": "https://www.lgresearch.ai/blog/view?seq=460" }, { "name": "Qwen2-Math", "org": "Alibaba", "country": "China", "date": "2024-08-08", "precision": "day", "category": "science", "open_weights": true, "params": "1.5B-72B", "note": "Math-specialised Qwen2 models that beat GPT-4o on MATH", "source": "https://qwenlm.github.io/blog/qwen2-math/" }, { "name": "Falcon Mamba 7B", "org": "TII", "country": "UAE", "date": "2024-08-12", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "First strong attention-free (pure Mamba) open 7B language model", "source": "https://huggingface.co/blog/falconmamba" }, { "name": "Grok-2", "org": "xAI", "country": "USA", "date": "2024-08-13", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "xAI frontier model; tested anonymously on LMSYS as sus-column-r", "source": "https://x.ai/news/grok-2" }, { "name": "Grok-2 mini", "org": "xAI", "country": "USA", "date": "2024-08-13", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Smaller, faster sibling of Grok-2", "source": "https://x.ai/news/grok-2" }, { "name": "Sarvam-2B", "org": "Sarvam AI", "country": "India", "date": "2024-08-13", "precision": "day", "category": "language", "open_weights": true, "params": "2B", "note": "Open small LLM for 10 Indic languages, launched with speech models", "source": "https://www.maginative.com/article/sarvam-ai-launches-open-weights-sarvam-2b-model-and-suite-of-ai-products-for-indian-languages/" }, { "name": "The AI Scientist", "org": "Sakana AI", "country": "Japan", "date": "2024-08-13", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "LLM agent automating ideation, experiments and paper writing end-to-end", "source": "https://sakana.ai/ai-scientist/" }, { "name": "Gen-3 Alpha Turbo", "org": "Runway", "country": "USA", "date": "2024-08-14", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Seven times faster, half-price variant of Gen-3 Alpha", "source": "https://runway.com/changelog" }, { "name": "Llama-3.1-Minitron 4B", "org": "NVIDIA", "country": "USA", "date": "2024-08-14", "precision": "day", "category": "language", "open_weights": true, "params": "4B", "note": "Pruned and distilled from Llama 3.1 8B", "source": "https://developer.nvidia.com/blog/how-to-prune-and-distill-llama-3-1-8b-to-an-nvidia-llama-3-1-minitron-4b-model/" }, { "name": "DeepSeek-Prover-V1.5", "org": "DeepSeek", "country": "China", "date": "2024-08-15", "precision": "day", "category": "science", "open_weights": true, "params": "7B", "note": "Lean 4 prover using proof-assistant-feedback RL and tree search", "source": "https://arxiv.org/abs/2408.08152" }, { "name": "Hermes 3", "org": "Nous Research", "country": "USA", "date": "2024-08-15", "precision": "day", "category": "language", "open_weights": true, "params": "8B, 70B, 405B", "note": "Neutrally aligned full-parameter fine-tunes of Llama 3.1 up to 405B", "source": "https://nousresearch.com/wp-content/uploads/2024/08/Hermes-3-Technical-Report.pdf" }, { "name": "xGen-MM v1.5 (BLIP-3)", "org": "Salesforce", "country": "USA", "date": "2024-08-16", "precision": "day", "category": "language", "open_weights": true, "params": "4B", "note": "Open multimodal models released with BLIP-3 report; v1 appeared May 2024", "source": "https://arxiv.org/abs/2408.08872" }, { "name": "Dream Machine 1.5", "org": "Luma AI", "country": "USA", "date": "2024-08-19", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Upgraded video model with better realism, motion and text rendering", "source": "https://www.techradar.com/computing/artificial-intelligence/dream-machine-15-catches-sora-and-other-rival-ai-video-makers-napping" }, { "name": "StormCast", "org": "NVIDIA", "country": "USA", "date": "2024-08-19", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Generative diffusion model emulating kilometre-scale storm-resolving weather forecasts", "source": "https://blogs.nvidia.com/blog/stormcast-generative-ai-weather-prediction/" }, { "name": "Phi-3.5-mini", "org": "Microsoft", "country": "USA", "date": "2024-08-20", "precision": "day", "category": "language", "open_weights": true, "params": "3.8B", "note": "Multilingual small model with 128K context", "source": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct" }, { "name": "Phi-3.5-MoE", "org": "Microsoft", "country": "USA", "date": "2024-08-20", "precision": "day", "category": "language", "open_weights": true, "params": "42B (6.6B active)", "note": "Microsoft's first mixture-of-experts Phi model", "source": "https://huggingface.co/microsoft/Phi-3.5-MoE-instruct" }, { "name": "Phi-3.5-vision", "org": "Microsoft", "country": "USA", "date": "2024-08-20", "precision": "day", "category": "language", "open_weights": true, "params": "4.2B", "note": "Small multi-frame image understanding model", "source": "https://huggingface.co/microsoft/Phi-3.5-vision-instruct" }, { "name": "Ideogram 2.0", "org": "Ideogram", "country": "Canada", "date": "2024-08-21", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Text-to-image model with strong typography and style presets", "source": "https://about.ideogram.ai/2.0" }, { "name": "Mistral-NeMo-Minitron 8B", "org": "NVIDIA", "country": "USA", "date": "2024-08-21", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Pruned, distilled Mistral NeMo leading 8B-class benchmarks", "source": "https://blogs.nvidia.com/blog/mistral-nemo-minitron-8b-small-language-model/" }, { "name": "Jamba 1.5 Large", "org": "AI21 Labs", "country": "Israel", "date": "2024-08-22", "precision": "day", "category": "language", "open_weights": true, "params": "398B (94B active)", "note": "Hybrid Transformer-Mamba MoE with 256K effective context", "source": "https://www.ai21.com/blog/announcing-jamba-model-family" }, { "name": "Jamba 1.5 Mini", "org": "AI21 Labs", "country": "Israel", "date": "2024-08-22", "precision": "day", "category": "language", "open_weights": true, "params": "52B (12B active)", "note": "Smaller hybrid SSM-Transformer long-context model", "source": "https://www.ai21.com/blog/announcing-jamba-model-family" }, { "name": "Sapiens", "org": "Meta", "country": "USA", "date": "2024-08-22", "precision": "day", "category": "vision", "open_weights": true, "params": "2B", "note": "Human-centric vision models for pose, segmentation, depth and normals", "source": "https://arxiv.org/abs/2408.12569" }, { "name": "Pharia-1-LLM-7B", "org": "Aleph Alpha", "country": "Germany", "date": "2024-08-26", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Transparent, EU AI Act-oriented model for German, French, Spanish", "source": "https://aleph-alpha.com/introducing-pharia-1-llm-transparent-and-compliant/" }, { "name": "CogVideoX-5B", "org": "Zhipu AI", "country": "China", "date": "2024-08-27", "precision": "day", "category": "video", "open_weights": true, "params": "5B", "note": "Larger open CogVideoX text-to-video model after the 2B release", "source": "https://github.com/zai-org/CogVideo" }, { "name": "GameNGen", "org": "Google", "country": "USA", "date": "2024-08-27", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Diffusion model simulating DOOM in real time as a neural game engine", "source": "https://arxiv.org/abs/2408.14837" }, { "name": "Gemini 1.5 Flash-8B", "org": "Google DeepMind", "country": "UK", "date": "2024-08-27", "precision": "day", "category": "multimodal", "open_weights": false, "params": "8B", "note": "Smallest Gemini 1.5; experimental in August, generally available October", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "Zamba2-1.2B", "org": "Zyphra", "country": "USA", "date": "2024-08-27", "precision": "day", "category": "language", "open_weights": true, "params": "1.2B", "note": "On-device hybrid SSM model (Zamba2-mini) under 700MB at 4-bit", "source": "https://www.zyphra.com/our-work" }, { "name": "Bielik-11B-v2", "org": "SpeakLeash", "country": "Poland", "date": "2024-08-28", "precision": "day", "category": "language", "open_weights": true, "params": "11B", "note": "Polish-focused open LLM trained with ACK Cyfronet AGH", "source": "https://pl.wikipedia.org/wiki/Bielik_(model_językowy)" }, { "name": "GLM-4-Plus", "org": "Zhipu AI", "country": "China", "date": "2024-08-29", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Zhipu flagship upgrade claimed near GPT-4o level", "source": "https://en.wikipedia.org/wiki/Zhipu_AI" }, { "name": "LTM-2-mini", "org": "Magic", "country": "USA", "date": "2024-08-29", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Model with a 100M-token context window aimed at code", "source": "https://magic.dev/blog/100m-token-context-windows" }, { "name": "Mini-Omni", "org": "Tsinghua University", "country": "China", "date": "2024-08-29", "precision": "day", "category": "audio", "open_weights": true, "params": "0.5B", "note": "Open end-to-end streaming speech model that talks while thinking", "source": "https://arxiv.org/abs/2408.16725" }, { "name": "Qwen2-VL", "org": "Alibaba", "country": "China", "date": "2024-08-29", "precision": "day", "category": "language", "open_weights": true, "params": "2B, 7B, 72B", "note": "VLM with dynamic resolution and 20-minute video understanding", "source": "https://qwenlm.github.io/blog/qwen2-vl/" }, { "name": "HelixFold3", "org": "Baidu", "country": "China", "date": "2024-08-30", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "AlphaFold3-style biomolecular structure predictor from PaddleHelix", "source": "https://arxiv.org/abs/2408.16975" }, { "name": "NV-Embed-v2", "org": "NVIDIA", "country": "USA", "date": "2024-08-30", "precision": "day", "category": "language", "open_weights": true, "params": "7.8B", "note": "Text embedding model ranked No. 1 on MTEB (72.31)", "source": "https://huggingface.co/nvidia/NV-Embed-v2" }, { "name": "video-01", "org": "MiniMax", "country": "China", "date": "2024-08-31", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Hailuo AI's breakout 720p text-to-video model", "source": "https://www.minimax.io/news/video-01" }, { "name": "GOT-OCR2.0", "org": "StepFun", "country": "China", "date": "2024-09-03", "precision": "day", "category": "vision", "open_weights": true, "params": "580M", "note": "Unified end-to-end 'OCR-2.0' model", "source": "https://arxiv.org/abs/2409.01704" }, { "name": "OLMoE", "org": "Allen Institute for AI", "country": "USA", "date": "2024-09-03", "precision": "day", "category": "language", "open_weights": true, "params": "7B (1B active)", "note": "Fully open mixture-of-experts model with open data and code", "source": "https://arxiv.org/abs/2409.02060" }, { "name": "AlphaProteo", "org": "Google DeepMind", "country": "UK", "date": "2024-09-05", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Designs novel high-affinity protein binders for target proteins", "source": "https://deepmind.google/discover/blog/alphaproteo-generates-novel-proteins-for-biology-and-health-research/" }, { "name": "DeepSeek-V2.5", "org": "DeepSeek", "country": "China", "date": "2024-09-05", "precision": "day", "category": "language", "open_weights": true, "params": "236B (21B active)", "note": "Merged DeepSeek-V2-Chat and Coder-V2 into one open model", "source": "https://api-docs.deepseek.com/news/news0905" }, { "name": "Harrison.rad.1", "org": "Harrison.ai", "country": "Australia", "date": "2024-09-05", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Radiology vision-language model for X-ray chat, findings and reports", "source": "https://harrison.ai/news/harrison-ai-launches-world-leading-ai-model-to-transform-healthcare/" }, { "name": "Hunyuan Turbo", "org": "Tencent", "country": "China", "date": "2024-09-05", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "MoE upgrade of Tencent's flagship with faster, half-price inference", "source": "https://zhidx.com/p/442250.html" }, { "name": "MiniCPM3-4B", "org": "OpenBMB", "country": "China", "date": "2024-09-05", "precision": "day", "category": "language", "open_weights": true, "params": "4B", "note": "4B model claimed comparable to 7B-9B models", "source": "https://huggingface.co/openbmb/MiniCPM3-4B" }, { "name": "Reflection 70B", "org": "HyperWrite", "country": "USA", "date": "2024-09-05", "precision": "day", "category": "language", "open_weights": true, "params": "70B", "note": "Hyped Llama fine-tune whose benchmark claims failed independent replication", "source": "https://huggingface.co/mattshumer/Reflection-Llama-3.1-70B" }, { "name": "xLAM", "org": "Salesforce", "country": "USA", "date": "2024-09-05", "precision": "day", "category": "agent", "open_weights": true, "params": "1B-141B", "note": "Large action model family (1B to 8x22B) for function calling", "source": "https://www.salesforce.com/blog/xlam-large-action-models/" }, { "name": "Yi-Coder", "org": "01.AI", "country": "China", "date": "2024-09-05", "precision": "day", "category": "code", "open_weights": true, "params": "1.5B, 9B", "note": "Open code models with 128K context across 52 languages", "source": "https://venturebeat.com/ai/yi-coder-the-open-source-ai-that-wants-to-be-your-coding-buddy" }, { "name": "Chai-1", "org": "Chai Discovery", "country": "USA", "date": "2024-09-09", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open multimodal biomolecular structure model rivaling AlphaFold3", "source": "https://www.chaidiscovery.com/blog/introducing-chai-1" }, { "name": "LLaMA-Omni", "org": "Chinese Academy of Sciences", "country": "China", "date": "2024-09-10", "precision": "day", "category": "audio", "open_weights": true, "params": "8B", "note": "Low-latency speech interaction model built on Llama-3.1-8B", "source": "https://arxiv.org/abs/2409.06666" }, { "name": "EVI 2", "org": "Hume AI", "country": "USA", "date": "2024-09-11", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Emotionally expressive voice-to-voice foundation model", "source": "https://www.hume.ai/blog/introducing-evi2" }, { "name": "Firefly Video Model", "org": "Adobe", "country": "USA", "date": "2024-09-11", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Commercially safe video model; previewed Sept 11, public beta Oct 14", "source": "https://blog.adobe.com/en/publish/2024/09/11/bringing-gen-ai-to-video-adobe-firefly-video-model-coming-soon" }, { "name": "Pixtral 12B", "org": "Mistral AI", "country": "France", "date": "2024-09-11", "precision": "day", "category": "language", "open_weights": true, "params": "12B", "note": "Mistral's first multimodal model; Apache 2.0 weights released via torrent", "source": "https://simonwillison.net/2024/Sep/11/pixtral/" }, { "name": "Solar Pro Preview", "org": "Upstage", "country": "South Korea", "date": "2024-09-11", "precision": "day", "category": "language", "open_weights": true, "params": "22B", "note": "22B model designed to run on a single GPU", "source": "https://www.upstage.ai/news/solar-pro-preview" }, { "name": "ALOHA Unleashed", "org": "Google DeepMind", "country": "UK", "date": "2024-09-12", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Bimanual robot policy learning dexterous tasks like tying shoelaces", "source": "https://deepmind.google/discover/blog/advances-in-robot-dexterity/" }, { "name": "DataGemma", "org": "Google", "country": "USA", "date": "2024-09-12", "precision": "day", "category": "language", "open_weights": true, "params": "27B", "note": "Gemma models grounded in Data Commons statistics to curb hallucination", "source": "https://blog.google/technology/ai/google-datagemma-ai-llm/" }, { "name": "DemoStart", "org": "Google DeepMind", "country": "UK", "date": "2024-09-12", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Simulation-trained RL for multi-fingered robot hand dexterity", "source": "https://deepmind.google/discover/blog/advances-in-robot-dexterity/" }, { "name": "o1-mini", "org": "OpenAI", "country": "USA", "date": "2024-09-12", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Cheaper, faster reasoning model strong at coding and math", "source": "https://openai.com/index/openai-o1-mini-advancing-cost-efficient-reasoning/" }, { "name": "o1-preview", "org": "OpenAI", "country": "USA", "date": "2024-09-12", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "First model trained with RL to reason via long hidden chain of thought", "source": "https://openai.com/index/introducing-openai-o1-preview/" }, { "name": "Seed-Music", "org": "ByteDance", "country": "China", "date": "2024-09-13", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Controllable vocal music generation and editing framework", "source": "https://arxiv.org/abs/2409.09214" }, { "name": "Playground v3", "org": "Playground", "country": "USA", "date": "2024-09-16", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "LLM-integrated text-to-image model strong at graphic design", "source": "https://arxiv.org/abs/2409.10695" }, { "name": "1X World Model", "org": "1X", "country": "Norway", "date": "2024-09-17", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Learned simulator predicting futures to evaluate humanoid robot policies", "source": "https://www.1x.tech/discover/1x-world-model" }, { "name": "Mistral Small 24.09", "org": "Mistral AI", "country": "France", "date": "2024-09-17", "precision": "day", "category": "language", "open_weights": true, "params": "22B", "note": "Open-weight 22B upgrade of Mistral Small", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "NVLM 1.0", "org": "NVIDIA", "country": "USA", "date": "2024-09-17", "precision": "day", "category": "language", "open_weights": true, "params": "72B", "note": "Frontier-class open multimodal LLMs rivaling GPT-4o on vision tasks", "source": "https://arxiv.org/abs/2409.11402" }, { "name": "OmniGen", "org": "BAAI", "country": "China", "date": "2024-09-17", "precision": "day", "category": "image", "open_weights": true, "params": "3.8B", "note": "Unified diffusion model for diverse image generation and editing tasks", "source": "https://arxiv.org/abs/2409.11340" }, { "name": "jina-embeddings-v3", "org": "Jina AI", "country": "Germany", "date": "2024-09-18", "precision": "day", "category": "language", "open_weights": true, "params": "570M", "note": "Multilingual embedding model with task-specific LoRA adapters", "source": "https://jina.ai/news/jina-embeddings-v3-a-frontier-multilingual-embedding-model/" }, { "name": "voyage-3", "org": "Voyage AI", "country": "USA", "date": "2024-09-18", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "General-purpose retrieval embedding model, released with voyage-3-lite", "source": "https://blog.voyageai.com/2024/09/18/voyage-3/" }, { "name": "Kling 1.5", "org": "Kuaishou", "country": "China", "date": "2024-09-19", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "1080p HD video generation plus motion brush", "source": "https://www.prnewswire.com/news-releases/kuaishou-technology-announces-third-quarter-2024-unaudited-financial-results-302311164.html" }, { "name": "Qwen2.5", "org": "Alibaba", "country": "China", "date": "2024-09-19", "precision": "day", "category": "language", "open_weights": true, "params": "72B", "note": "Open family 0.5B-72B trained on 18T tokens", "source": "https://qwenlm.github.io/blog/qwen2.5/" }, { "name": "Qwen2.5-Coder", "org": "Alibaba", "country": "China", "date": "2024-09-19", "precision": "day", "category": "code", "open_weights": true, "params": "7B", "note": "Code-specialised Qwen2.5 models trained on 5.5T tokens", "source": "https://arxiv.org/abs/2409.12186" }, { "name": "Qwen2.5-Math", "org": "Alibaba", "country": "China", "date": "2024-09-19", "precision": "day", "category": "science", "open_weights": true, "params": "72B", "note": "Math models with chain-of-thought and tool-integrated reasoning", "source": "https://qwenlm.github.io/blog/qwen2.5-math/" }, { "name": "Tongyi Wanxiang (video)", "org": "Alibaba", "country": "China", "date": "2024-09-19", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Alibaba's first text-to-video model, unveiled at Apsara Conference 2024", "source": "https://www.alibabacloud.com/blog/alibaba-cloud-unveils-new-ai-models-and-revamped-infrastructure-for-ai-computing_601622" }, { "name": "Prithvi WxC", "org": "IBM", "country": "USA", "date": "2024-09-20", "precision": "day", "category": "science", "open_weights": true, "params": "2.3B", "note": "IBM-NASA foundation model for weather and climate", "source": "https://arxiv.org/abs/2409.13598" }, { "name": "TeleChat2-115B", "org": "China Telecom", "country": "China", "date": "2024-09-20", "precision": "day", "category": "language", "open_weights": true, "params": "115B", "note": "Open 100B-plus model trained entirely on domestic Chinese compute", "source": "https://github.com/Tele-AI/TeleChat2" }, { "name": "Llama-3.1-Nemotron-51B", "org": "NVIDIA", "country": "USA", "date": "2024-09-23", "precision": "day", "category": "language", "open_weights": true, "params": "51B", "note": "Architecture-searched Llama 3.1 70B derivative fitting one H100", "source": "https://developer.nvidia.com/blog/advancing-the-accuracy-efficiency-frontier-with-llama-3-1-nemotron-51b/" }, { "name": "Doubao PixelDance", "org": "ByteDance", "country": "China", "date": "2024-09-24", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Doubao video model for multi-shot, multi-subject interaction", "source": "https://kr-asia.com/bytedance-enters-ai-video-race-with-doubaos-pixeldance-and-seaweed" }, { "name": "Doubao Seaweed", "org": "ByteDance", "country": "China", "date": "2024-09-24", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Second Doubao video generation model unveiled alongside PixelDance", "source": "https://kr-asia.com/bytedance-enters-ai-video-race-with-doubaos-pixeldance-and-seaweed" }, { "name": "Gemini-1.5-Pro-002", "org": "Google DeepMind", "country": "UK", "date": "2024-09-24", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Updated Gemini 1.5 Pro and Flash (-002) with large price cuts", "source": "https://developers.googleblog.com/en/updated-gemini-models-reduced-15-pro-pricing-increased-rate-limits-and-more/" }, { "name": "Llama 3.2 (1B, 3B)", "org": "Meta", "country": "USA", "date": "2024-09-25", "precision": "day", "category": "language", "open_weights": true, "params": "1B, 3B", "note": "Lightweight text-only Llamas for on-device and edge use", "source": "https://huggingface.co/blog/llama32" }, { "name": "Llama 3.2 Vision (11B, 90B)", "org": "Meta", "country": "USA", "date": "2024-09-25", "precision": "day", "category": "language", "open_weights": true, "params": "11B, 90B", "note": "First open Llama models with image understanding", "source": "https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/" }, { "name": "Molmo", "org": "Allen Institute for AI", "country": "USA", "date": "2024-09-25", "precision": "day", "category": "language", "open_weights": true, "params": "72B", "note": "Open VLM family with open PixMo data and pointing ability", "source": "https://allenai.org/blog/molmo" }, { "name": "omni-moderation-latest", "org": "OpenAI", "country": "USA", "date": "2024-09-26", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "GPT-4o-based moderation model handling both text and images", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "Emu3", "org": "BAAI", "country": "China", "date": "2024-09-27", "precision": "day", "category": "multimodal", "open_weights": true, "params": "8B", "note": "Next-token prediction alone for image/video generation and understanding", "source": "https://arxiv.org/abs/2409.18869" }, { "name": "YOLO11", "org": "Ultralytics", "country": "USA", "date": "2024-09-27", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Latest YOLO real-time detection, segmentation and pose models", "source": "https://docs.ultralytics.com/models/yolo11/" }, { "name": "SAM 2.1", "org": "Meta", "country": "USA", "date": "2024-09-29", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Improved Segment Anything 2 checkpoints plus developer suite", "source": "https://github.com/facebookresearch/sam2" }, { "name": "HPT", "org": "MIT", "country": "USA", "date": "2024-09-30", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Heterogeneous Pre-trained Transformers pooling data across robot embodiments", "source": "https://arxiv.org/abs/2409.20537" }, { "name": "Liquid Foundation Models (LFM)", "org": "Liquid AI", "country": "USA", "date": "2024-09-30", "precision": "day", "category": "language", "open_weights": false, "params": "40.3B (12B active)", "note": "Non-transformer LFM-1B, 3B and 40B MoE models from MIT spinout", "source": "https://www.liquid.ai/liquid-foundation-models" }, { "name": "MM1.5", "org": "Apple", "country": "USA", "date": "2024-09-30", "precision": "day", "category": "language", "open_weights": false, "params": "30B", "note": "Apple multimodal LLM family 1B-30B focused on text-rich images", "source": "https://arxiv.org/abs/2409.20566" }, { "name": "Takane", "org": "Fujitsu", "country": "Japan", "date": "2024-09-30", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Japanese enterprise LLM co-developed with Cohere", "source": "https://www.fujitsu.com/global/about/resources/news/press-releases/2024/0930-01.html" }, { "name": "Teuken-7B", "org": "OpenGPT-X (Fraunhofer)", "country": "Germany", "date": "2024-09-30", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Open model covering all 24 official EU languages", "source": "https://arxiv.org/abs/2410.03730" }, { "name": "GPT-4o Realtime", "org": "OpenAI", "country": "USA", "date": "2024-10-01", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-to-speech gpt-4o-realtime-preview via DevDay Realtime API", "source": "https://openai.com/index/introducing-the-realtime-api/" }, { "name": "Pika 1.5", "org": "Pika", "country": "USA", "date": "2024-10-01", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Video model update debuting viral Pikaffects physics effects", "source": "https://x.com/pika_labs/status/1841143349576941863" }, { "name": "Whisper large-v3-turbo", "org": "OpenAI", "country": "USA", "date": "2024-10-01", "precision": "day", "category": "audio", "open_weights": true, "params": "809M", "note": "Pruned Whisper with 4 decoder layers, about 8x faster", "source": "https://huggingface.co/openai/whisper-large-v3-turbo" }, { "name": "Depth Pro", "org": "Apple", "country": "USA", "date": "2024-10-02", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Sharp zero-shot metric monocular depth in under a second", "source": "https://arxiv.org/abs/2410.02073" }, { "name": "FLUX1.1 [pro]", "org": "Black Forest Labs", "country": "Germany", "date": "2024-10-02", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Six times faster successor that topped the Artificial Analysis image arena", "source": "https://bfl.ai/blog/24-10-02-flux" }, { "name": "Gemma 2 JPN", "org": "Google DeepMind", "country": "UK", "date": "2024-10-03", "precision": "day", "category": "language", "open_weights": true, "params": "2B", "note": "Japanese-tuned Gemma 2 2B", "source": "https://ai.google.dev/gemma/docs/releases" }, { "name": "Movie Gen", "org": "Meta", "country": "USA", "date": "2024-10-04", "precision": "day", "category": "video", "open_weights": false, "params": "30B", "note": "30B video and 13B audio models for generation and editing", "source": "https://ai.meta.com/blog/movie-gen-media-foundation-models-generative-ai-video/" }, { "name": "Inflection 3.0", "org": "Inflection AI", "country": "USA", "date": "2024-10-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Enterprise-grade LLM powering Inflection for Enterprise on Intel Gaudi 3", "source": "https://www.inflection.ai/blog/enterprise" }, { "name": "Aria", "org": "Rhymes AI", "country": "USA", "date": "2024-10-08", "precision": "day", "category": "language", "open_weights": true, "params": "25.3B (3.9B active)", "note": "Open multimodal-native mixture-of-experts model", "source": "https://arxiv.org/abs/2410.05993" }, { "name": "GR-2", "org": "ByteDance", "country": "China", "date": "2024-10-08", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Generative video-language-action robot model pretrained on 38M videos", "source": "https://arxiv.org/abs/2410.06158" }, { "name": "Pyramid Flow", "org": "Peking University", "country": "China", "date": "2024-10-08", "precision": "day", "category": "video", "open_weights": true, "params": "2B", "note": "Open pyramidal flow-matching video model built with Kuaishou", "source": "https://arxiv.org/abs/2410.05954" }, { "name": "F5-TTS", "org": "Shanghai Jiao Tong University", "country": "China", "date": "2024-10-09", "precision": "day", "category": "audio", "open_weights": true, "params": "336M", "note": "Open flow-matching non-autoregressive zero-shot TTS", "source": "https://arxiv.org/abs/2410.06885" }, { "name": "Palmyra X 004", "org": "Writer", "country": "USA", "date": "2024-10-09", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Enterprise LLM with tool calling and actions", "source": "https://writer.com/engineering/actions-with-palmyra-x-004/" }, { "name": "RDT-1B", "org": "Tsinghua University", "country": "China", "date": "2024-10-10", "precision": "day", "category": "robotics", "open_weights": true, "params": "1.2B", "note": "Diffusion foundation model for bimanual robot manipulation", "source": "https://arxiv.org/abs/2410.07864" }, { "name": "Baichuan-Omni", "org": "Baichuan", "country": "China", "date": "2024-10-11", "precision": "day", "category": "multimodal", "open_weights": false, "params": "7B", "note": "Omni-modal 7B model handling text, image, video and audio", "source": "https://arxiv.org/abs/2410.08565" }, { "name": "INTELLECT-1", "org": "Prime Intellect", "country": "USA", "date": "2024-10-11", "precision": "day", "category": "language", "open_weights": true, "params": "10B", "note": "First globally distributed 10B training run; weights released Nov 29", "source": "https://www.primeintellect.ai/blog/intellect-1" }, { "name": "SANA", "org": "NVIDIA", "country": "USA", "date": "2024-10-14", "precision": "day", "category": "image", "open_weights": true, "params": "1.6B", "note": "Linear-attention diffusion model generating up to 4K images efficiently", "source": "https://arxiv.org/abs/2410.10629" }, { "name": "Zamba2-7B", "org": "Zyphra", "country": "USA", "date": "2024-10-14", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Hybrid SSM-attention 7B claimed best-in-class small model", "source": "https://www.zyphra.com/post/zamba2-7b" }, { "name": "Llama-3.1-Nemotron-70B-Instruct", "org": "NVIDIA", "country": "USA", "date": "2024-10-15", "precision": "day", "category": "language", "open_weights": true, "params": "70B", "note": "RLHF-tuned Llama that topped Arena Hard and AlpacaEval at release", "source": "https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF/commits/main" }, { "name": "Ministral 3B", "org": "Mistral AI", "country": "France", "date": "2024-10-16", "precision": "day", "category": "language", "open_weights": false, "params": "3B", "note": "Edge-focused model in the 'les Ministraux' family", "source": "https://mistral.ai/news/ministraux" }, { "name": "Ministral 8B", "org": "Mistral AI", "country": "France", "date": "2024-10-16", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Edge model with interleaved sliding-window attention; research-license weights", "source": "https://mistral.ai/news/ministraux" }, { "name": "OMat24", "org": "Meta", "country": "USA", "date": "2024-10-16", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open inorganic materials dataset and pretrained interatomic potential models", "source": "https://arxiv.org/abs/2410.12771" }, { "name": "Yi-Lightning", "org": "01.AI", "country": "China", "date": "2024-10-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Cheap MoE flagship ranked near top of LMSYS; trained on ~2,000 GPUs", "source": "https://pandaily.com/01-ai-releases-new-flagship-model-yi-lightning" }, { "name": "Fluid", "org": "Google DeepMind", "country": "UK", "date": "2024-10-17", "precision": "day", "category": "image", "open_weights": false, "params": "10.5B", "note": "Continuous-token autoregressive text-to-image model, with MIT", "source": "https://arxiv.org/abs/2410.13863" }, { "name": "gpt-4o-audio-preview", "org": "OpenAI", "country": "USA", "date": "2024-10-17", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Audio input and output in the Chat Completions API", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "Janus", "org": "DeepSeek", "country": "China", "date": "2024-10-17", "precision": "day", "category": "multimodal", "open_weights": true, "params": "1.3B", "note": "Unified understanding and image generation with decoupled visual encoders", "source": "https://arxiv.org/abs/2410.13848" }, { "name": "Allegro", "org": "Rhymes AI", "country": "USA", "date": "2024-10-20", "precision": "day", "category": "video", "open_weights": true, "params": "2.8B", "note": "Open commercial-level text-to-video model", "source": "https://arxiv.org/abs/2410.15458" }, { "name": "Granite 3.0", "org": "IBM", "country": "USA", "date": "2024-10-21", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Apache-2.0 enterprise LLMs with disclosed training data", "source": "https://www.ibm.com/new/announcements/ibm-granite-3-0-open-state-of-the-art-enterprise-models" }, { "name": "Moonshine", "org": "Useful Sensors", "country": "USA", "date": "2024-10-21", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Edge speech recognition models faster than Whisper for live transcription", "source": "https://arxiv.org/abs/2410.15608" }, { "name": "Act-One", "org": "Runway", "country": "USA", "date": "2024-10-22", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Animates characters from a single driving performance video", "source": "https://runwayml.com/research/introducing-act-one" }, { "name": "Claude 3.5 Haiku", "org": "Anthropic", "country": "USA", "date": "2024-10-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Fast model matching Claude 3 Opus on many evals; shipped Nov 4", "source": "https://www.anthropic.com/news/3-5-models-and-computer-use" }, { "name": "Claude 3.5 Sonnet (new)", "org": "Anthropic", "country": "USA", "date": "2024-10-22", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Upgraded Sonnet debuting computer use, first frontier model operating computers", "source": "https://www.anthropic.com/news/3-5-models-and-computer-use" }, { "name": "Embed 3 (multimodal)", "org": "Cohere", "country": "Canada", "date": "2024-10-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Embedding model unifying text and image search", "source": "https://cohere.com/blog/multimodal-embed-3" }, { "name": "Mochi 1", "org": "Genmo", "country": "USA", "date": "2024-10-22", "precision": "day", "category": "video", "open_weights": true, "params": "10B", "note": "Largest open-weight video generation model at release, Apache 2.0", "source": "https://github.com/genmoai/mochi" }, { "name": "Stable Diffusion 3.5 Large", "org": "Stability AI", "country": "UK", "date": "2024-10-22", "precision": "day", "category": "image", "open_weights": true, "params": "8.1B", "note": "Open MMDiT flagship following the SD3 Medium backlash", "source": "https://stability.ai/news/introducing-stable-diffusion-3-5" }, { "name": "Stable Diffusion 3.5 Large Turbo", "org": "Stability AI", "country": "UK", "date": "2024-10-22", "precision": "day", "category": "image", "open_weights": true, "params": "8.1B", "note": "Distilled four-step version of SD 3.5 Large", "source": "https://stability.ai/news/introducing-stable-diffusion-3-5" }, { "name": "Aya Expanse", "org": "Cohere", "country": "Canada", "date": "2024-10-24", "precision": "day", "category": "language", "open_weights": true, "params": "32B", "note": "Open multilingual models (8B, 32B) covering 23 languages", "source": "https://cohere.com/blog/aya-expanse-connecting-our-world" }, { "name": "Sarvam-1", "org": "Sarvam AI", "country": "India", "date": "2024-10-24", "precision": "day", "category": "language", "open_weights": true, "params": "2B", "note": "2B model optimised for 10 Indic languages", "source": "https://www.sarvam.ai/blogs/sarvam-1" }, { "name": "Spark 4.0 Turbo", "org": "iFlytek", "country": "China", "date": "2024-10-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Math ability claimed above GPT-4o; multilingual model debut", "source": "https://cn.chinadaily.com.cn/a/202410/24/WS671a0ea6a310b59111d9fbe6.html" }, { "name": "YandexGPT 4", "org": "Yandex", "country": "Russia", "date": "2024-10-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Pro and Lite models with improved reasoning and 32K context", "source": "https://yandex.com/company/news/24-10-2024" }, { "name": "AutoGLM", "org": "Zhipu AI", "country": "China", "date": "2024-10-25", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Phone and web GUI agent executing tasks from natural language", "source": "https://xiao9905.github.io/AutoGLM/" }, { "name": "GLM-4-Voice", "org": "Zhipu AI", "country": "China", "date": "2024-10-25", "precision": "day", "category": "audio", "open_weights": true, "params": "9B", "note": "End-to-end Chinese-English speech LLM that adjusts tone and dialect", "source": "https://github.com/THUDM/GLM-4-Voice" }, { "name": "Stable Diffusion 3.5 Medium", "org": "Stability AI", "country": "UK", "date": "2024-10-29", "precision": "day", "category": "image", "open_weights": true, "params": "2.5B", "note": "Consumer-GPU MMDiT-X variant of SD 3.5", "source": "https://stability.ai/news/introducing-stable-diffusion-3-5" }, { "name": "EMMA", "org": "Waymo", "country": "USA", "date": "2024-10-30", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "End-to-end multimodal autonomous driving model built on Gemini", "source": "https://waymo.com/blog/2024/10/introducing-emma" }, { "name": "Recraft V3", "org": "Recraft", "country": "UK", "date": "2024-10-30", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "'red_panda' model that topped image arena; strong at design and text", "source": "https://www.recraft.ai/blog/recraft-introduces-a-revolutionary-ai-model-that-thinks-in-design-language" }, { "name": "Universal-2", "org": "AssemblyAI", "country": "USA", "date": "2024-10-30", "precision": "day", "category": "audio", "open_weights": false, "params": "600M", "note": "Speech-to-text model improving proper nouns, formatting and timestamps", "source": "https://www.assemblyai.com/research/universal-2" }, { "name": "AMD OLMo", "org": "AMD", "country": "USA", "date": "2024-10-31", "precision": "day", "category": "language", "open_weights": true, "params": "1B", "note": "AMD's first fully open 1B models, trained on Instinct MI250 GPUs", "source": "https://www.amd.com/en/developer/resources/technical-articles/introducing-the-first-amd-1b-language-model.html" }, { "name": "Oasis", "org": "Decart", "country": "Israel", "date": "2024-10-31", "precision": "day", "category": "video", "open_weights": true, "params": "500M", "note": "Real-time playable AI-generated Minecraft-like world, built with Etched", "source": "https://oasis-model.github.io/" }, { "name": "Sparsh", "org": "Meta", "country": "USA", "date": "2024-10-31", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Self-supervised touch representations for vision-based tactile sensors", "source": "https://arxiv.org/abs/2410.24090" }, { "name": "π0", "org": "Physical Intelligence", "country": "USA", "date": "2024-10-31", "precision": "day", "category": "robotics", "open_weights": false, "params": "3.3B", "note": "Generalist vision-language-action flow model for dexterous robot control", "source": "https://arxiv.org/abs/2410.24164" }, { "name": "Big Sleep", "org": "Google", "country": "USA", "date": "2024-11-01", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Project Zero/DeepMind agent found first unknown exploitable bug (SQLite)", "source": "https://projectzero.google/2024/10/from-naptime-to-big-sleep.html" }, { "name": "SmolLM2", "org": "Hugging Face", "country": "USA", "date": "2024-11-01", "precision": "day", "category": "language", "open_weights": true, "params": "1.7B", "note": "Improved 135M, 360M and 1.7B small open models", "source": "https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct/commits/main" }, { "name": "hertz-dev", "org": "Standard Intelligence", "country": "USA", "date": "2024-11-03", "precision": "day", "category": "audio", "open_weights": true, "params": "8.5B", "note": "Open base model for full-duplex conversational audio", "source": "https://si.inc/posts/hertz-dev" }, { "name": "Magentic-One", "org": "Microsoft", "country": "USA", "date": "2024-11-04", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Generalist multi-agent system orchestrating web, file and coding agents", "source": "https://www.microsoft.com/en-us/research/articles/magentic-one-a-generalist-multi-agent-system-for-solving-complex-tasks/" }, { "name": "Hunyuan-Large", "org": "Tencent", "country": "China", "date": "2024-11-05", "precision": "day", "category": "language", "open_weights": true, "params": "389B (52B active)", "note": "Largest open Transformer mixture-of-experts model at release", "source": "https://arxiv.org/abs/2411.02265" }, { "name": "Hunyuan3D-1.0", "org": "Tencent", "country": "China", "date": "2024-11-05", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open text- and image-to-3D generation framework", "source": "https://github.com/Tencent/Hunyuan3D-1" }, { "name": "FLUX1.1 [pro] Ultra", "org": "Black Forest Labs", "country": "Germany", "date": "2024-11-06", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "4-megapixel Ultra mode plus photorealistic Raw mode", "source": "https://bfl.ai/blog/24-11-06-ultra" }, { "name": "Mistral Moderation", "org": "Mistral AI", "country": "France", "date": "2024-11-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Multilingual LLM content-moderation classifier behind Le Chat", "source": "https://mistral.ai/news/mistral-moderation" }, { "name": "OpenCoder", "org": "INF Technology", "country": "China", "date": "2024-11-07", "precision": "day", "category": "code", "open_weights": true, "params": "8B", "note": "Fully open code LLM releasing data and training pipeline", "source": "https://arxiv.org/abs/2411.04905" }, { "name": "CogVideoX1.5", "org": "Zhipu AI", "country": "China", "date": "2024-11-08", "precision": "day", "category": "video", "open_weights": true, "params": "5B", "note": "Upgraded open CogVideoX with longer, higher-resolution clips", "source": "https://github.com/zai-org/CogVideo" }, { "name": "SeedEdit", "org": "ByteDance", "country": "China", "date": "2024-11-11", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Diffusion-based instruction-driven image editing model", "source": "https://seed.bytedance.com/en/seededit" }, { "name": "Qwen2.5-Coder-32B-Instruct", "org": "Alibaba", "country": "China", "date": "2024-11-12", "precision": "day", "category": "code", "open_weights": true, "params": "32B", "note": "Open code model competitive with GPT-4o on coding benchmarks", "source": "https://qwenlm.github.io/blog/qwen2.5-coder-family/" }, { "name": "JanusFlow", "org": "DeepSeek", "country": "China", "date": "2024-11-13", "precision": "day", "category": "multimodal", "open_weights": true, "params": "1.3B", "note": "Unifies autoregression with rectified flow for understanding and generation", "source": "https://arxiv.org/abs/2411.07975" }, { "name": "Vidu 1.5", "org": "ShengShu", "country": "China", "date": "2024-11-13", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Multi-subject consistent video generation from reference images", "source": "https://petapixel.com/2024/11/14/vidu-is-a-chinese-ai-video-generator-that-lets-you-combine-still-images-into-one-video/" }, { "name": "Athene-V2", "org": "Nexusflow", "country": "USA", "date": "2024-11-14", "precision": "day", "category": "language", "open_weights": true, "params": "72B", "note": "Qwen2.5-72B post-trained chat and agent models", "source": "https://huggingface.co/Nexusflow/Athene-V2-Chat" }, { "name": "Gemini-Exp-1114", "org": "Google DeepMind", "country": "UK", "date": "2024-11-14", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Experimental Gemini that jumped to #1 on Chatbot Arena", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "LLaVA-CoT", "org": "Peking University", "country": "China", "date": "2024-11-15", "precision": "day", "category": "reasoning", "open_weights": true, "params": "11B", "note": "Vision-language model doing structured multistage reasoning", "source": "https://arxiv.org/abs/2411.10440" }, { "name": "Qwen2.5-Turbo", "org": "Alibaba", "country": "China", "date": "2024-11-15", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Extends Qwen context to 1M tokens with faster inference", "source": "https://qwenlm.github.io/blog/qwen2.5-turbo/" }, { "name": "Boltz-1", "org": "MIT", "country": "USA", "date": "2024-11-16", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Fully open biomolecular structure model at AlphaFold 3 level", "source": "https://jclinic.mit.edu/boltz-1/" }, { "name": "k0-math", "org": "Moonshot AI", "country": "China", "date": "2024-11-16", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Math reasoning model claimed to rival o1 on math benchmarks", "source": "https://web.archive.org/web/20250124103542/https://www.globaltimes.cn/page/202411/1323248.shtml" }, { "name": "Mistral Large 24.11", "org": "Mistral AI", "country": "France", "date": "2024-11-18", "precision": "day", "category": "language", "open_weights": true, "params": "123B", "note": "Update improving long context, function calling and system prompts", "source": "https://huggingface.co/mistralai/Mistral-Large-Instruct-2411" }, { "name": "Pixtral Large", "org": "Mistral AI", "country": "France", "date": "2024-11-18", "precision": "day", "category": "language", "open_weights": true, "params": "124B", "note": "Frontier-class open-weight multimodal model built on Mistral Large 2", "source": "https://mistral.ai/news/pixtral-large" }, { "name": "Suno v4", "org": "Suno", "country": "USA", "date": "2024-11-19", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music model with cleaner audio, sharper lyrics and dynamic structures", "source": "https://suno.com/blog/v4" }, { "name": "AlphaQubit", "org": "Google DeepMind", "country": "UK", "date": "2024-11-20", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Transformer decoder identifying errors in quantum computers", "source": "https://blog.google/technology/google-deepmind/alphaqubit-quantum-error-correction/" }, { "name": "DeepSeek-R1-Lite-Preview", "org": "DeepSeek", "country": "China", "date": "2024-11-20", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "DeepSeek's first reasoning model, o1-preview-level on AIME", "source": "https://api-docs.deepseek.com/news/news1120" }, { "name": "Hymba", "org": "NVIDIA", "country": "USA", "date": "2024-11-20", "precision": "day", "category": "language", "open_weights": true, "params": "1.5B", "note": "Small model fusing attention and Mamba heads in parallel", "source": "https://arxiv.org/abs/2411.13676" }, { "name": "AIMv2", "org": "Apple", "country": "USA", "date": "2024-11-21", "precision": "day", "category": "vision", "open_weights": true, "params": "2.7B", "note": "Autoregressively pre-trained multimodal vision encoders", "source": "https://arxiv.org/abs/2411.14402" }, { "name": "FLUX.1 Tools", "org": "Black Forest Labs", "country": "Germany", "date": "2024-11-21", "precision": "day", "category": "image", "open_weights": true, "params": "12B", "note": "Fill, Depth, Canny and Redux models for control and editing", "source": "https://bfl.ai/blog/24-11-21-tools" }, { "name": "Gemini-Exp-1121", "org": "Google DeepMind", "country": "UK", "date": "2024-11-21", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Experimental Gemini that reclaimed #1 on Chatbot Arena", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "LTX-Video", "org": "Lightricks", "country": "Israel", "date": "2024-11-21", "precision": "day", "category": "video", "open_weights": true, "params": "2B", "note": "Open DiT video model generating faster than real time", "source": "https://github.com/Lightricks/LTX-Video" }, { "name": "Marco-o1", "org": "Alibaba", "country": "China", "date": "2024-11-21", "precision": "day", "category": "reasoning", "open_weights": true, "params": "7B", "note": "Open reasoning model using MCTS for open-ended problems", "source": "https://arxiv.org/abs/2411.14405" }, { "name": "Samsung Gauss2", "org": "Samsung", "country": "South Korea", "date": "2024-11-21", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Second-generation Samsung model in Compact, Balanced, Supreme sizes", "source": "https://news.samsung.com/global/samsung-electronics-hosts-samsung-developer-conference-korea-2024-unveils-its-improved-gen-ai-model" }, { "name": "Tülu 3", "org": "Allen Institute for AI", "country": "USA", "date": "2024-11-21", "precision": "day", "category": "language", "open_weights": true, "params": "8B / 70B", "note": "Fully open post-training recipe introducing RL with verifiable rewards", "source": "https://allenai.org/blog/tulu-3" }, { "name": "Fish Speech 1.5", "org": "Fish Audio", "country": "USA", "date": "2024-11-24", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open multilingual text-to-speech model", "source": "https://huggingface.co/fishaudio/fish-speech-1.5" }, { "name": "Frames", "org": "Runway", "country": "USA", "date": "2024-11-25", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Runway image model emphasising stylistic control", "source": "https://runway.com/research/introducing-frames" }, { "name": "Fugatto", "org": "NVIDIA", "country": "USA", "date": "2024-11-25", "precision": "day", "category": "audio", "open_weights": false, "params": "2.5B", "note": "Generative audio model for music, voice and sound", "source": "https://research.nvidia.com/publication/2024-11_fugatto-1-foundational-generative-audio-transformer-opus-1" }, { "name": "Luma Photon", "org": "Luma AI", "country": "USA", "date": "2024-11-25", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Luma's image model, with cheaper Photon Flash variant", "source": "https://lumalabs.ai/photon" }, { "name": "OLMo 2", "org": "Allen Institute for AI", "country": "USA", "date": "2024-11-26", "precision": "day", "category": "language", "open_weights": true, "params": "7B / 13B", "note": "Fully open 7B and 13B models rivaling Llama 3.1 8B", "source": "https://allenai.org/blog/olmo2" }, { "name": "SmolVLM", "org": "Hugging Face", "country": "USA", "date": "2024-11-26", "precision": "day", "category": "language", "open_weights": true, "params": "2B", "note": "Compact open vision-language model for on-device use", "source": "https://huggingface.co/blog/smolvlm" }, { "name": "Skywork o1 Open", "org": "Kunlun Tech", "country": "China", "date": "2024-11-27", "precision": "day", "category": "reasoning", "open_weights": true, "params": "8B", "note": "Early open o1-style reasoning models with process reward models", "source": "https://huggingface.co/Skywork/Skywork-o1-Open-Llama-3.1-8B" }, { "name": "QwQ-32B-Preview", "org": "Alibaba", "country": "China", "date": "2024-11-28", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32.5B", "note": "First open-weight o1-style reasoning model", "source": "https://qwenlm.github.io/blog/qwq-32b-preview/" }, { "name": "EuroLLM-9B", "org": "EuroLLM consortium", "country": "Portugal", "date": "2024-12-02", "precision": "day", "category": "language", "open_weights": true, "params": "9B", "note": "Open LLM covering all 24 official EU languages", "source": "https://huggingface.co/blog/eurollm-team/eurollm-9b" }, { "name": "Rerank 3.5", "org": "Cohere", "country": "Canada", "date": "2024-12-02", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Reranking model with improved reasoning and multilingual search", "source": "https://cohere.com/blog/rerank-3pt5" }, { "name": "TRELLIS", "org": "Microsoft", "country": "USA", "date": "2024-12-02", "precision": "day", "category": "image", "open_weights": true, "params": "2B", "note": "Structured 3D latents for high-quality 3D asset generation", "source": "https://arxiv.org/abs/2412.01506" }, { "name": "World Labs world generation (preview)", "org": "World Labs", "country": "USA", "date": "2024-12-02", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Turns a single image into explorable 3D worlds in the browser", "source": "https://www.worldlabs.ai/blog/generating-worlds" }, { "name": "Amazon Nova Canvas", "org": "Amazon", "country": "USA", "date": "2024-12-03", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Amazon's image generation and editing model", "source": "https://aws.amazon.com/about-aws/whats-new/2024/12/amazon-nova-foundation-models-bedrock" }, { "name": "Amazon Nova Lite", "org": "Amazon", "country": "USA", "date": "2024-12-03", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Low-cost model taking text, image and video input", "source": "https://aws.amazon.com/blogs/aws/introducing-amazon-nova-frontier-intelligence-and-industry-leading-price-performance/" }, { "name": "Amazon Nova Micro", "org": "Amazon", "country": "USA", "date": "2024-12-03", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Text-only, lowest-latency model of Amazon's first frontier family", "source": "https://aws.amazon.com/blogs/aws/introducing-amazon-nova-frontier-intelligence-and-industry-leading-price-performance/" }, { "name": "Amazon Nova Pro", "org": "Amazon", "country": "USA", "date": "2024-12-03", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Amazon's most capable Nova model at launch", "source": "https://aws.amazon.com/blogs/aws/introducing-amazon-nova-frontier-intelligence-and-industry-leading-price-performance/" }, { "name": "Amazon Nova Reel", "org": "Amazon", "country": "USA", "date": "2024-12-03", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Amazon's video generation model", "source": "https://aws.amazon.com/about-aws/whats-new/2024/12/amazon-nova-foundation-models-bedrock" }, { "name": "HunyuanVideo", "org": "Tencent", "country": "China", "date": "2024-12-03", "precision": "day", "category": "video", "open_weights": true, "params": "13B", "note": "Largest open video generation model at release", "source": "https://arxiv.org/abs/2412.03603" }, { "name": "ESM C", "org": "EvolutionaryScale", "country": "USA", "date": "2024-12-04", "precision": "day", "category": "science", "open_weights": true, "params": "300M / 600M / 6B", "note": "ESM Cambrian protein language models succeeding ESM2", "source": "https://www.evolutionaryscale.ai/blog/esm-cambrian" }, { "name": "GenCast", "org": "Google DeepMind", "country": "UK", "date": "2024-12-04", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Diffusion ensemble weather model beating ECMWF ENS up to 15 days", "source": "https://deepmind.google/discover/blog/gencast-predicts-weather-and-the-risks-of-extreme-conditions-with-sota-accuracy/" }, { "name": "Genie 2", "org": "Google DeepMind", "country": "UK", "date": "2024-12-04", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Foundation world model generating playable 3D environments from one image", "source": "https://deepmind.google/discover/blog/genie-2-a-large-scale-foundation-world-model/" }, { "name": "Infinity", "org": "ByteDance", "country": "China", "date": "2024-12-05", "precision": "day", "category": "image", "open_weights": true, "params": "2B", "note": "Bitwise autoregressive text-to-image model rivaling diffusion models", "source": "https://arxiv.org/abs/2412.04431" }, { "name": "InternVL2.5", "org": "Shanghai AI Lab", "country": "China", "date": "2024-12-05", "precision": "day", "category": "language", "open_weights": true, "params": "1B-78B", "note": "First open multimodal LLM to exceed 70% on MMMU", "source": "https://internvl.github.io/blog/2024-12-05-InternVL-2.5/" }, { "name": "o1", "org": "OpenAI", "country": "USA", "date": "2024-12-05", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Full o1 release with image input, replacing o1-preview", "source": "https://en.wikipedia.org/wiki/OpenAI_o1" }, { "name": "o1 pro mode", "org": "OpenAI", "country": "USA", "date": "2024-12-05", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Higher-compute o1 in the new $200/month ChatGPT Pro", "source": "https://openai.com/index/introducing-chatgpt-pro/" }, { "name": "PaliGemma 2", "org": "Google DeepMind", "country": "UK", "date": "2024-12-05", "precision": "day", "category": "language", "open_weights": true, "params": "3B / 10B / 28B", "note": "Open VLMs pairing SigLIP with Gemma 2", "source": "https://developers.googleblog.com/en/introducing-paligemma-2-powerful-vision-language-models-simple-fine-tuning/" }, { "name": "Pleias 1.0", "org": "PleIAs", "country": "France", "date": "2024-12-05", "precision": "day", "category": "language", "open_weights": true, "params": "350M-3B", "note": "Small models trained only on open, permissively licensed data", "source": "https://huggingface.co/PleIAs/Pleias-1.2b-Preview" }, { "name": "Solar Pro", "org": "Upstage", "country": "South Korea", "date": "2024-12-05", "precision": "day", "category": "language", "open_weights": false, "params": "22B", "note": "Billed as the most intelligent LLM on a single GPU", "source": "https://www.upstage.ai/blog" }, { "name": "Gemini-Exp-1206", "org": "Google DeepMind", "country": "UK", "date": "2024-12-06", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Experimental Gemini topping Chatbot Arena on Gemini's first anniversary", "source": "https://simonwillison.net/2024/Dec/6/gemini-exp-1206/" }, { "name": "Llama 3.3 70B", "org": "Meta", "country": "USA", "date": "2024-12-06", "precision": "day", "category": "language", "open_weights": true, "params": "70B", "note": "70B model matching Llama 3.1 405B at far lower cost", "source": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct" }, { "name": "Aurora", "org": "xAI", "country": "USA", "date": "2024-12-09", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Autoregressive MoE image generator powering Grok on X", "source": "https://x.ai/news/grok-image-generation-release" }, { "name": "EXAONE 3.5", "org": "LG AI Research", "country": "South Korea", "date": "2024-12-09", "precision": "day", "category": "language", "open_weights": true, "params": "2.4B / 7.8B / 32B", "note": "Long-context bilingual Korean-English open model family", "source": "https://arxiv.org/abs/2412.04862" }, { "name": "Sora Turbo", "org": "OpenAI", "country": "USA", "date": "2024-12-09", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Faster Sora launched publicly to ChatGPT Plus and Pro users", "source": "https://openai.com/index/sora-is-here/" }, { "name": "DeepSeek-V2.5-1210", "org": "DeepSeek", "country": "China", "date": "2024-12-10", "precision": "day", "category": "language", "open_weights": true, "params": "236B (21B active)", "note": "Final V2.5 update improving math, coding and writing", "source": "https://api-docs.deepseek.com/news/news1210" }, { "name": "Gemini 2.0 Flash", "org": "Google DeepMind", "country": "UK", "date": "2024-12-11", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Agentic-era Gemini with native image/audio output and Live API", "source": "https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/" }, { "name": "Gemini Deep Research", "org": "Google", "country": "USA", "date": "2024-12-11", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Agent that browses the web to compile multi-page research reports", "source": "https://blog.google/products/gemini/google-gemini-deep-research/" }, { "name": "Jules", "org": "Google", "country": "USA", "date": "2024-12-11", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Experimental Gemini 2.0-powered asynchronous coding agent for GitHub", "source": "https://developers.googleblog.com/en/the-next-chapter-of-the-gemini-era-for-developers/" }, { "name": "Project Mariner", "org": "Google DeepMind", "country": "UK", "date": "2024-12-11", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Gemini 2.0 research prototype agent that operates the Chrome browser", "source": "https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/" }, { "name": "T-Pro", "org": "T-Bank", "country": "Russia", "date": "2024-12-11", "precision": "day", "category": "language", "open_weights": true, "params": "32B", "note": "Open Russian-language model on Qwen2.5, released with 7B T-Lite", "source": "https://habr.com/ru/companies/tbank/articles/865582/" }, { "name": "Byte Latent Transformer", "org": "Meta", "country": "USA", "date": "2024-12-12", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Tokenizer-free byte-level model with dynamic patches", "source": "https://arxiv.org/abs/2412.09871" }, { "name": "Grok-2-1212", "org": "xAI", "country": "USA", "date": "2024-12-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Faster Grok-2 update released as Grok became free on X", "source": "https://x.ai/news/grok-1212" }, { "name": "Large Concept Model", "org": "Meta", "country": "USA", "date": "2024-12-12", "precision": "day", "category": "language", "open_weights": false, "params": "7B", "note": "Language modeling over sentence-level concepts instead of tokens", "source": "https://arxiv.org/abs/2412.08821" }, { "name": "Meta Motivo", "org": "Meta", "country": "USA", "date": "2024-12-12", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Behavioral foundation model for zero-shot humanoid control", "source": "https://ai.meta.com/blog/meta-fair-updates-agents-robustness-safety-architecture/" }, { "name": "Meta Video Seal", "org": "Meta", "country": "USA", "date": "2024-12-12", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Open neural video watermarking model robust to edits", "source": "https://ai.meta.com/blog/meta-fair-updates-agents-robustness-safety-architecture/" }, { "name": "Phi-4", "org": "Microsoft", "country": "USA", "date": "2024-12-12", "precision": "day", "category": "language", "open_weights": true, "params": "14B", "note": "Synthetic-data-heavy small model excelling at math reasoning", "source": "https://arxiv.org/abs/2412.08905" }, { "name": "Command R7B", "org": "Cohere", "country": "Canada", "date": "2024-12-13", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Smallest, fastest Command R model for RAG and agents", "source": "https://cohere.com/blog/command-r7b" }, { "name": "CosyVoice 2", "org": "Alibaba", "country": "China", "date": "2024-12-13", "precision": "day", "category": "audio", "open_weights": true, "params": "0.5B", "note": "Streaming zero-shot TTS with about 150ms latency", "source": "https://arxiv.org/abs/2412.10117" }, { "name": "DeepSeek-VL2", "org": "DeepSeek", "country": "China", "date": "2024-12-13", "precision": "day", "category": "language", "open_weights": true, "params": "27B (4.5B active)", "note": "MoE vision-language models for OCR, documents and grounding", "source": "https://arxiv.org/abs/2412.10302" }, { "name": "GigaChat-20B-A3B", "org": "Sber", "country": "Russia", "date": "2024-12-13", "precision": "day", "category": "language", "open_weights": true, "params": "20B (3B active)", "note": "Sber's first open-weight GigaChat MoE model", "source": "https://huggingface.co/api/models?author=ai-sage&sort=createdAt&direction=-1&limit=60" }, { "name": "Kandinsky 4.0", "org": "Sber", "country": "Russia", "date": "2024-12-13", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Text-to-video family (Flash open-sourced) plus video-to-audio model", "source": "https://github.com/ai-forever/Kandinsky-4" }, { "name": "Pika 2.0", "org": "Pika", "country": "USA", "date": "2024-12-13", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Adds Scene Ingredients for inserting user-supplied characters and objects", "source": "https://www.linkedin.com/posts/pika-labs_today-we-launched-our-pika-20-model-superior-activity-7273465246919380994-8R40" }, { "name": "Veo 2", "org": "Google DeepMind", "country": "UK", "date": "2024-12-16", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "4K-capable video model with better physics; led human preference tests", "source": "https://blog.google/technology/google-labs/video-image-generation-update-december-2024/" }, { "name": "Falcon 3", "org": "TII", "country": "UAE", "date": "2024-12-17", "precision": "day", "category": "language", "open_weights": true, "params": "1B-10B", "note": "Open family of 1B-10B models trained on 14T tokens", "source": "https://huggingface.co/blog/falcon3" }, { "name": "FastVLM", "org": "Apple", "country": "USA", "date": "2024-12-17", "precision": "day", "category": "language", "open_weights": true, "params": "0.5B-7B", "note": "Efficient hybrid vision encoding for on-device VLMs", "source": "https://arxiv.org/abs/2412.13303" }, { "name": "Eleven Flash v2.5", "org": "ElevenLabs", "country": "USA", "date": "2024-12-18", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Ultra-low-latency (about 75ms) TTS for conversational agents", "source": "https://elevenlabs.io/blog/meet-flash" }, { "name": "Granite 3.1", "org": "IBM", "country": "USA", "date": "2024-12-18", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Granite update with 128K context and new embedding models", "source": "https://www.ibm.com/new/announcements/ibm-granite-3-1-powerful-performance-long-context-and-more" }, { "name": "Typhoon 2", "org": "SCB 10X", "country": "Thailand", "date": "2024-12-18", "precision": "day", "category": "language", "open_weights": true, "params": "1B-70B", "note": "Open Thai text, vision and audio models built on Llama and Qwen", "source": "https://arxiv.org/abs/2412.13702" }, { "name": "Gemini 2.0 Flash Thinking Experimental", "org": "Google DeepMind", "country": "UK", "date": "2024-12-19", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Google's first thinking model with visible reasoning", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "Kling 1.6", "org": "Kuaishou", "country": "China", "date": "2024-12-19", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Better prompt adherence and motion quality", "source": "https://ir.kuaishou.com/news-releases/news-release-details/kuaishou-kling-ai-unveils-multi-image-reference-feature-further/" }, { "name": "ModernBERT", "org": "Answer.AI", "country": "USA", "date": "2024-12-19", "precision": "day", "category": "language", "open_weights": true, "params": "395M", "note": "First major BERT-style encoder refresh in years; 8K context, with LightOn", "source": "https://huggingface.co/blog/modernbert" }, { "name": "CogAgent-9B-20241220", "org": "Zhipu AI", "country": "China", "date": "2024-12-20", "precision": "day", "category": "agent", "open_weights": true, "params": "9B", "note": "Open screenshot-only GUI agent VLM underpinning GLM-PC", "source": "https://github.com/THUDM/CogAgent" }, { "name": "o3", "org": "OpenAI", "country": "USA", "date": "2024-12-20", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Announced with record ARC-AGI and FrontierMath scores; released April 2025", "source": "https://techcrunch.com/2024/12/20/openai-announces-new-o3-model/" }, { "name": "o3-mini", "org": "OpenAI", "country": "USA", "date": "2024-12-20", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Distilled o3 announced Dec 2024; released January 31, 2025", "source": "https://en.wikipedia.org/wiki/OpenAI_o3" }, { "name": "OCTAVE", "org": "Hume AI", "country": "USA", "date": "2024-12-23", "precision": "day", "category": "audio", "open_weights": false, "params": "3B", "note": "Speech-language model generating voices and personalities from prompts", "source": "https://www.hume.ai/blog/introducing-octave" }, { "name": "QVQ-72B-Preview", "org": "Alibaba", "country": "China", "date": "2024-12-24", "precision": "day", "category": "reasoning", "open_weights": true, "params": "72B", "note": "Open visual reasoning model built on Qwen2-VL-72B", "source": "https://qwenlm.github.io/blog/qvq-72b-preview/" }, { "name": "DeepSeek-V3", "org": "DeepSeek", "country": "China", "date": "2024-12-25", "precision": "day", "category": "language", "open_weights": true, "params": "671B (37B active)", "note": "Frontier-level open MoE trained on 2.788M H800 GPU hours", "source": "https://simonwillison.net/2024/Dec/25/deepseek-v3/" }, { "name": "Kokoro-82M", "org": "hexgrad", "country": "Unknown", "date": "2024-12-25", "precision": "day", "category": "audio", "open_weights": true, "params": "82M", "note": "Tiny open TTS model that topped the TTS Arena", "source": "https://huggingface.co/hexgrad/Kokoro-82M" }, { "name": "Cosmos World Foundation Models", "org": "NVIDIA", "country": "USA", "date": "2025-01-06", "precision": "day", "category": "video", "open_weights": true, "params": "4B-14B", "note": "Open physics-aware world models for robotics and driving, unveiled at CES", "source": "https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-world-foundation-model-platform-to-accelerate-physical-ai-development" }, { "name": "METAGENE-1", "org": "Prime Intellect", "country": "USA", "date": "2025-01-06", "precision": "day", "category": "science", "open_weights": true, "params": "7B", "note": "Metagenomic foundation model for pathogen detection", "source": "https://www.primeintellect.ai/blog/metagene" }, { "name": "Tiangong 4.0", "org": "Kunlun Tech", "country": "China", "date": "2025-01-06", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Skywork 4.0 o1 reasoning and 4o voice versions released free", "source": "https://www.aibase.com/news/14480" }, { "name": "voyage-3-large", "org": "Voyage AI", "country": "USA", "date": "2025-01-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "State-of-the-art general-purpose text embedding model", "source": "https://blog.voyageai.com/2025/01/07/voyage-3-large/" }, { "name": "rStar-Math", "org": "Microsoft", "country": "USA", "date": "2025-01-08", "precision": "day", "category": "reasoning", "open_weights": false, "params": "7B", "note": "Small models reach o1-level math via MCTS self-evolution", "source": "https://arxiv.org/abs/2501.04519" }, { "name": "Stable Point Aware 3D (SPAR3D)", "org": "Stability AI", "country": "UK", "date": "2025-01-08", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Real-time single-image to editable 3D object generation", "source": "https://stability.ai/news-updates/stable-point-aware-3d" }, { "name": "Sky-T1-32B-Preview", "org": "NovaSky (UC Berkeley)", "country": "USA", "date": "2025-01-10", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32B", "note": "Open reasoning model trained for under $450", "source": "https://novasky-ai.github.io/posts/sky-t1/" }, { "name": "Codestral 25.01", "org": "Mistral AI", "country": "France", "date": "2025-01-13", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Faster Codestral upgrade leading fill-in-the-middle coding benchmarks", "source": "https://mistral.ai/news/codestral-2501" }, { "name": "Helium 1", "org": "Kyutai", "country": "France", "date": "2025-01-13", "precision": "day", "category": "language", "open_weights": true, "params": "2B", "note": "Multilingual edge LLM; preview January, full release April 2025", "source": "https://huggingface.co/api/models/kyutai/helium-1-preview-2b" }, { "name": "MiniCPM-o 2.6", "org": "OpenBMB", "country": "China", "date": "2025-01-13", "precision": "day", "category": "multimodal", "open_weights": true, "params": "8B", "note": "GPT-4o-level vision, speech and live streaming on device", "source": "https://huggingface.co/openbmb/MiniCPM-o-2_6" }, { "name": "InternLM3-8B-Instruct", "org": "Shanghai AI Lab", "country": "China", "date": "2025-01-15", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Open 8B trained on only 4T tokens with deep-thinking mode", "source": "https://huggingface.co/internlm/internlm3-8b-instruct" }, { "name": "MiniMax-Text-01", "org": "MiniMax", "country": "China", "date": "2025-01-15", "precision": "day", "category": "language", "open_weights": true, "params": "456B (45.9B active)", "note": "Lightning-attention MoE with 4M-token context window", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "MiniMax-VL-01", "org": "MiniMax", "country": "China", "date": "2025-01-15", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Vision-language sibling of MiniMax-Text-01", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "Ray2", "org": "Luma AI", "country": "USA", "date": "2025-01-15", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Video model trained with 10x the compute of Ray1", "source": "https://the-decoder.com/luma-launches-ray2-its-latest-ai-video-model-scaled-to-10x-compute/" }, { "name": "Spark X1", "org": "iFlytek", "country": "China", "date": "2025-01-15", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Deep-reasoning model trained fully on domestic Chinese compute", "source": "https://en.tmtpost.com/news/7477906" }, { "name": "Vidu 2.0", "org": "Shengshu", "country": "China", "date": "2025-01-15", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Faster, cheaper video generation", "source": "https://kr-asia.com/ai-video-generation-just-got-faster-and-cheaper-with-shengshus-vidu-2-0" }, { "name": "π0-FAST", "org": "Physical Intelligence", "country": "USA", "date": "2025-01-16", "precision": "day", "category": "robotics", "open_weights": true, "params": "3B", "note": "Autoregressive VLA using FAST frequency-space action tokenizer", "source": "https://arxiv.org/abs/2501.09747" }, { "name": "GPT-4b micro", "org": "OpenAI", "country": "USA", "date": "2025-01-17", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Protein-engineering model built with Retro Biosciences for longevity", "source": "https://www.technologyreview.com/2025/01/17/1110086/openai-has-created-an-ai-model-for-longevity-science/" }, { "name": "INTELLECT-MATH", "org": "Prime Intellect", "country": "USA", "date": "2025-01-17", "precision": "day", "category": "reasoning", "open_weights": true, "params": "7B", "note": "Math reasoning model trained with RL", "source": "https://huggingface.co/PrimeIntellect/INTELLECT-MATH" }, { "name": "T2A-01-HD", "org": "MiniMax", "country": "China", "date": "2025-01-17", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "TTS with seconds-long voice cloning across 17 languages", "source": "https://www.minimax.cn/news/t2a-01-hd-%E5%8F%91%E5%B8%83" }, { "name": "DeepSeek-R1", "org": "DeepSeek", "country": "China", "date": "2025-01-20", "precision": "day", "category": "reasoning", "open_weights": true, "params": "671B (37B active)", "note": "MIT-licensed reasoning model matching o1; triggered global market shock", "source": "https://api-docs.deepseek.com/news/news250120" }, { "name": "DeepSeek-R1-Distill", "org": "DeepSeek", "country": "China", "date": "2025-01-20", "precision": "day", "category": "reasoning", "open_weights": true, "params": "1.5B-70B", "note": "Six dense Qwen/Llama models distilled from R1 reasoning traces", "source": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B" }, { "name": "DeepSeek-R1-Zero", "org": "DeepSeek", "country": "China", "date": "2025-01-20", "precision": "day", "category": "reasoning", "open_weights": true, "params": "671B (37B active)", "note": "Reasoning emerged from pure RL without supervised fine-tuning", "source": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero" }, { "name": "Doubao Realtime Voice Model", "org": "ByteDance", "country": "China", "date": "2025-01-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "End-to-end real-time speech conversation model", "source": "https://seed.bytedance.com/blog/doubao-realtime-voice-model-is-available-upon-release-high-eq-and-iq" }, { "name": "Kimi k1.5", "org": "Moonshot AI", "country": "China", "date": "2025-01-20", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Multimodal RL reasoning model claiming o1-level math", "source": "https://github.com/MoonshotAI/Kimi-k1.5" }, { "name": "LFM-7B", "org": "Liquid AI", "country": "USA", "date": "2025-01-20", "precision": "day", "category": "language", "open_weights": false, "params": "7B", "note": "Efficient multilingual non-transformer Liquid model", "source": "https://www.liquid.ai/blog/introducing-lfm-7b-setting-new-standards-for-efficient-language-models" }, { "name": "Gemini 2.0 Flash Thinking Experimental 01-21", "org": "Google DeepMind", "country": "UK", "date": "2025-01-21", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Updated thinking model with 1M context; topped Chatbot Arena", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "Hunyuan3D 2.0", "org": "Tencent", "country": "China", "date": "2025-01-21", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open high-resolution image-to-3D asset generation system", "source": "https://github.com/Tencent-Hunyuan/Hunyuan3D-2" }, { "name": "UI-TARS", "org": "ByteDance", "country": "China", "date": "2025-01-21", "precision": "day", "category": "agent", "open_weights": true, "params": "2B-72B", "note": "Native end-to-end GUI agent model", "source": "https://arxiv.org/abs/2501.12326" }, { "name": "Video Depth Anything", "org": "ByteDance", "country": "China", "date": "2025-01-21", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Consistent depth estimation for super-long videos", "source": "https://arxiv.org/abs/2501.12375" }, { "name": "Doubao-1.5-pro", "org": "ByteDance", "country": "China", "date": "2025-01-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Sparse MoE flagship priced far below GPT-4o", "source": "https://seed.bytedance.com/zh/special/doubao_1_5_pro" }, { "name": "Baichuan-M1", "org": "Baichuan", "country": "China", "date": "2025-01-23", "precision": "day", "category": "language", "open_weights": true, "params": "14B", "note": "Open LLM trained from scratch for medicine", "source": "https://huggingface.co/api/models/baichuan-inc/Baichuan-M1-14B-Instruct" }, { "name": "Computer-Using Agent (CUA)", "org": "OpenAI", "country": "USA", "date": "2025-01-23", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "GPT-4o-based model powering the Operator browser agent", "source": "https://openai.com/index/computer-using-agent/" }, { "name": "SmolVLM-256M / SmolVLM-500M", "org": "Hugging Face", "country": "USA", "date": "2025-01-23", "precision": "day", "category": "language", "open_weights": true, "params": "256M, 500M", "note": "World's smallest vision-language models at release", "source": "https://huggingface.co/blog/smolervlm" }, { "name": "Baichuan-Omni-1.5", "org": "Baichuan", "country": "China", "date": "2025-01-26", "precision": "day", "category": "multimodal", "open_weights": true, "params": "11B", "note": "Open omni-modal model with end-to-end speech", "source": "https://arxiv.org/abs/2501.15368" }, { "name": "Qwen2.5-VL", "org": "Alibaba", "country": "China", "date": "2025-01-26", "precision": "day", "category": "language", "open_weights": true, "params": "3B-72B", "note": "Vision-language models with agentic computer and phone control", "source": "https://qwenlm.github.io/blog/qwen2.5-vl/" }, { "name": "YuE", "org": "HKUST & M-A-P", "country": "China", "date": "2025-01-26", "precision": "day", "category": "audio", "open_weights": true, "params": "7B", "note": "Open lyrics-to-full-song music generation model", "source": "https://github.com/multimodal-art-projection/YuE/tree/YuE-v1" }, { "name": "Janus-Pro", "org": "DeepSeek", "country": "China", "date": "2025-01-27", "precision": "day", "category": "image", "open_weights": true, "params": "1B, 7B", "note": "Unified multimodal understanding and image generation", "source": "https://arxiv.org/abs/2501.17811" }, { "name": "Pika 2.1", "org": "Pika", "country": "USA", "date": "2025-01-27", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "1080p video generation with sharper detail and more realistic motion", "source": "https://www.linkedin.com/posts/pika-labs_pika-21-is-here-the-details-are-crystal-activity-7289689170954891265-CkzI" }, { "name": "Qwen2.5-1M", "org": "Alibaba", "country": "China", "date": "2025-01-27", "precision": "day", "category": "language", "open_weights": true, "params": "7B, 14B", "note": "Open models supporting 1M-token context", "source": "https://qwenlm.github.io/blog/qwen2.5-1m/" }, { "name": "Qwen2.5-Max", "org": "Alibaba", "country": "China", "date": "2025-01-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Large MoE flagship claimed to beat DeepSeek-V3", "source": "https://qwenlm.github.io/blog/qwen2.5-max/" }, { "name": "Mistral Small 3", "org": "Mistral AI", "country": "France", "date": "2025-01-30", "precision": "day", "category": "language", "open_weights": true, "params": "24B", "note": "Apache-licensed latency-optimized 24B model", "source": "https://mistral.ai/news/mistral-small-3" }, { "name": "TinySwallow-1.5B", "org": "Sakana AI", "country": "Japan", "date": "2025-01-30", "precision": "day", "category": "language", "open_weights": true, "params": "1.5B", "note": "Small Japanese model distilled with new TAID method", "source": "https://sakana.ai/taid-jp/" }, { "name": "Tülu 3 405B", "org": "Allen Institute for AI", "country": "USA", "date": "2025-01-30", "precision": "day", "category": "language", "open_weights": true, "params": "405B", "note": "Largest fully open post-trained model, using RLVR", "source": "https://allenai.org/blog/tulu-3-405B" }, { "name": "s1-32B", "org": "Stanford University", "country": "USA", "date": "2025-01-31", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32B", "note": "Test-time scaling via budget forcing from 1,000 examples", "source": "https://arxiv.org/abs/2501.19393" }, { "name": "ALLaM-7B-Instruct-preview", "org": "SDAIA", "country": "Saudi Arabia", "date": "2025-02-01", "precision": "month", "category": "language", "open_weights": true, "params": "7B", "note": "Open-weight preview of Saudi Arabic-English national LLM", "source": "https://huggingface.co/api/models/humain-ai/ALLaM-7B-Instruct-preview" }, { "name": "Ideogram 2a", "org": "Ideogram", "country": "Canada", "date": "2025-02-01", "precision": "month", "category": "image", "open_weights": false, "params": "", "note": "Faster, cheaper Ideogram model for graphic design and photography", "source": "https://en.wikipedia.org/wiki/Ideogram_(text-to-image_model)" }, { "name": "Krutrim-2", "org": "Krutrim", "country": "India", "date": "2025-02-01", "precision": "month", "category": "language", "open_weights": true, "params": "12B", "note": "Indic-focused open LLM from Ola's AI unit", "source": "https://huggingface.co/krutrim-ai-labs/Krutrim-2-instruct" }, { "name": "PLaMo 2", "org": "Preferred Networks", "country": "Japan", "date": "2025-02-01", "precision": "month", "category": "language", "open_weights": true, "params": "1B, 8B", "note": "Samba-based hybrid Japanese language models", "source": "https://huggingface.co/api/models/pfnet/plamo-2-8b" }, { "name": "Deep Research", "org": "OpenAI", "country": "USA", "date": "2025-02-02", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "o3-based agent that autonomously researches the web", "source": "https://techcrunch.com/2025/02/02/openai-unveils-a-new-chatgpt-agent-for-deep-research/" }, { "name": "OmniHuman-1", "org": "ByteDance", "country": "China", "date": "2025-02-03", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Single-image, audio-driven realistic human video generation", "source": "https://arxiv.org/abs/2502.01061" }, { "name": "AlphaGeometry2", "org": "Google DeepMind", "country": "UK", "date": "2025-02-05", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Gold-medalist-level performance on olympiad geometry problems", "source": "https://arxiv.org/abs/2502.03544" }, { "name": "BFS-Prover", "org": "ByteDance", "country": "China", "date": "2025-02-05", "precision": "day", "category": "science", "open_weights": true, "params": "7B", "note": "State-of-the-art Lean 4 formal theorem prover", "source": "https://arxiv.org/abs/2502.03438" }, { "name": "Gemini 2.0 Flash-Lite", "org": "Google DeepMind", "country": "UK", "date": "2025-02-05", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "First Flash-Lite model, optimized for cost efficiency", "source": "https://developers.googleblog.com/en/gemini-2-family-expands/" }, { "name": "Gemini 2.0 Pro Experimental", "org": "Google DeepMind", "country": "UK", "date": "2025-02-05", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Strongest Gemini 2.0 for coding, 2M-token context", "source": "https://blog.google/technology/google-deepmind/gemini-model-updates-february-2025/" }, { "name": "Hibiki", "org": "Kyutai", "country": "France", "date": "2025-02-05", "precision": "day", "category": "audio", "open_weights": true, "params": "2B", "note": "Simultaneous on-device speech-to-speech translation preserving voice", "source": "https://arxiv.org/abs/2502.03382" }, { "name": "Brain2Qwerty", "org": "Meta", "country": "USA", "date": "2025-02-07", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Decodes typed sentences from non-invasive MEG/EEG signals", "source": "https://www.marktechpost.com/2025/02/09/meta-ai-introduces-brain2qwerty-a-new-deep-learning-model-for-decoding-sentences-from-brain-activity-with-eeg-or-meg-while-participants-typed-briefly-memorized-sentences-on-a-qwerty-keyboard/" }, { "name": "Goku", "org": "ByteDance", "country": "China", "date": "2025-02-07", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Rectified-flow joint image and video generation models", "source": "https://arxiv.org/abs/2502.04896" }, { "name": "VideoWorld", "org": "ByteDance", "country": "China", "date": "2025-02-10", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open research model learning world knowledge from video alone", "source": "https://seed.bytedance.com/blog/seed-research-latest-breakthrough-in-video-generation-learning-to-understand-the-world-through-vision-alone-now-open-source" }, { "name": "Zonos-v0.1", "org": "Zyphra", "country": "USA", "date": "2025-02-10", "precision": "day", "category": "audio", "open_weights": true, "params": "1.6B", "note": "Expressive open TTS with high-fidelity voice cloning", "source": "https://www.zyphra.com/post/beta-release-of-zonos-v0-1" }, { "name": "Sonar", "org": "Perplexity", "country": "USA", "date": "2025-02-11", "precision": "day", "category": "language", "open_weights": false, "params": "70B", "note": "In-house Llama 3.3 70B-based search model tuned for factuality", "source": "https://the-decoder.com/perplexity-ai-launches-new-ultra-fast-ai-search-model-sonar/" }, { "name": "T2V-01-Director", "org": "MiniMax", "country": "China", "date": "2025-02-11", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Hailuo T2V/I2V-01-Director models with natural-language camera control", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "OpenThinker-32B", "org": "Open Thoughts", "country": "USA", "date": "2025-02-12", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32B", "note": "Open-data reasoning model rivaling R1-Distill-32B", "source": "https://huggingface.co/open-thoughts/OpenThinker-32B" }, { "name": "DeepHermes 3 Preview", "org": "Nous Research", "country": "USA", "date": "2025-02-13", "precision": "day", "category": "reasoning", "open_weights": true, "params": "8B", "note": "Toggles between normal replies and long chain-of-thought", "source": "https://huggingface.co/NousResearch/DeepHermes-3-Llama-3-8B-Preview" }, { "name": "LLaDA", "org": "Renmin University of China & Ant Group", "country": "China", "date": "2025-02-14", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Diffusion language model rivaling autoregressive LLaMA3 8B", "source": "https://arxiv.org/abs/2502.09992" }, { "name": "Image-01", "org": "MiniMax", "country": "China", "date": "2025-02-15", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "MiniMax's first text-to-image model", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "Grok 3", "org": "xAI", "country": "USA", "date": "2025-02-17", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Colossus-trained flagship with Think reasoning mode and DeepSearch", "source": "https://x.ai/news/grok-3" }, { "name": "Grok 3 mini", "org": "xAI", "country": "USA", "date": "2025-02-17", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Cost-efficient reasoning variant of Grok 3", "source": "https://x.ai/news/grok-3" }, { "name": "Mistral Saba", "org": "Mistral AI", "country": "France", "date": "2025-02-17", "precision": "day", "category": "language", "open_weights": false, "params": "24B", "note": "Regional model for Arabic and South Asian languages", "source": "https://mistral.ai/news/mistral-saba" }, { "name": "Step-Audio", "org": "StepFun", "country": "China", "date": "2025-02-17", "precision": "day", "category": "audio", "open_weights": true, "params": "130B", "note": "Open 130B speech understanding and generation model", "source": "https://arxiv.org/abs/2502.11946" }, { "name": "Step-Video-T2V", "org": "StepFun", "country": "China", "date": "2025-02-17", "precision": "day", "category": "video", "open_weights": true, "params": "30B", "note": "30B-parameter open text-to-video model", "source": "https://github.com/stepfun-ai/Step-Video-T2V" }, { "name": "Magma", "org": "Microsoft", "country": "USA", "date": "2025-02-18", "precision": "day", "category": "agent", "open_weights": true, "params": "8B", "note": "Foundation model for multimodal UI and robotic agents", "source": "https://arxiv.org/abs/2502.13130" }, { "name": "R1 1776", "org": "Perplexity", "country": "USA", "date": "2025-02-18", "precision": "day", "category": "reasoning", "open_weights": true, "params": "671B (37B active)", "note": "DeepSeek-R1 post-trained to remove Chinese censorship", "source": "https://the-decoder.com/perplexity-ai-removes-chinese-censorship-from-deepseek-r1/" }, { "name": "SkyReels V1", "org": "Kunlun Tech", "country": "China", "date": "2025-02-18", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open human-centric video model built on HunyuanVideo", "source": "https://github.com/SkyworkAI/SkyReels-V1" }, { "name": "YOLOv12", "org": "University at Buffalo & UCAS", "country": "USA", "date": "2025-02-18", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Attention-centric real-time object detector", "source": "https://arxiv.org/abs/2502.12524" }, { "name": "AI co-scientist", "org": "Google", "country": "USA", "date": "2025-02-19", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Gemini 2.0 multi-agent system generating research hypotheses", "source": "https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/" }, { "name": "Evo 2", "org": "Arc Institute", "country": "USA", "date": "2025-02-19", "precision": "day", "category": "science", "open_weights": true, "params": "40B", "note": "Genome model spanning all domains of life, built with NVIDIA", "source": "https://arcinstitute.org/news/evo2" }, { "name": "Muse (WHAM)", "org": "Microsoft", "country": "USA", "date": "2025-02-19", "precision": "day", "category": "games", "open_weights": true, "params": "", "note": "Generative world and human action model for gameplay ideation", "source": "https://www.microsoft.com/en-us/research/blog/introducing-muse-our-first-generative-ai-model-designed-for-gameplay-ideation/" }, { "name": "PaliGemma 2 mix", "org": "Google DeepMind", "country": "UK", "date": "2025-02-19", "precision": "day", "category": "language", "open_weights": true, "params": "3B-28B", "note": "Vision-language models tuned for many tasks out of the box", "source": "https://developers.googleblog.com/en/introducing-paligemma-2-mix/" }, { "name": "BioEmu-1", "org": "Microsoft", "country": "USA", "date": "2025-02-20", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Generates protein structural ensembles at thousands per hour", "source": "https://www.microsoft.com/en-us/research/blog/exploring-the-structural-changes-driving-protein-function-with-bioemu-1/" }, { "name": "Helix", "org": "Figure AI", "country": "USA", "date": "2025-02-20", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "VLA controlling a humanoid's full upper body", "source": "https://www.figure.ai/news/helix" }, { "name": "SigLIP 2", "org": "Google DeepMind", "country": "UK", "date": "2025-02-20", "precision": "day", "category": "vision", "open_weights": true, "params": "up to 1B", "note": "Multilingual vision-language encoders with dense features", "source": "https://arxiv.org/abs/2502.14786" }, { "name": "SmolVLM2", "org": "Hugging Face", "country": "USA", "date": "2025-02-20", "precision": "day", "category": "language", "open_weights": true, "params": "256M-2.2B", "note": "Small video-understanding models for every device", "source": "https://huggingface.co/blog/smolvlm2" }, { "name": "Moonlight", "org": "Moonshot AI", "country": "China", "date": "2025-02-22", "precision": "day", "category": "language", "open_weights": true, "params": "16B (3B active)", "note": "MoE model proving Muon optimizer scales to LLM training", "source": "https://huggingface.co/api/models/moonshotai/Moonlight-16B-A3B" }, { "name": "Claude 3.7 Sonnet", "org": "Anthropic", "country": "USA", "date": "2025-02-24", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "First hybrid reasoning model; launched alongside Claude Code", "source": "https://www.anthropic.com/news/claude-3-7-sonnet" }, { "name": "PLLuM", "org": "NASK", "country": "Poland", "date": "2025-02-24", "precision": "day", "category": "language", "open_weights": true, "params": "8B-70B", "note": "Polish government-backed LLM family from NASK-led consortium", "source": "https://pl.wikipedia.org/wiki/PLLuM" }, { "name": "AIFS Single v1.0", "org": "ECMWF", "country": "UK", "date": "2025-02-25", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "First operational machine-learning weather forecast at ECMWF", "source": "https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational" }, { "name": "olmOCR", "org": "Allen Institute for AI", "country": "USA", "date": "2025-02-25", "precision": "day", "category": "vision", "open_weights": true, "params": "7B", "note": "Open VLM toolkit for high-quality PDF text extraction", "source": "https://allenai.org/blog/olmocr" }, { "name": "QwQ-Max-Preview", "org": "Alibaba", "country": "China", "date": "2025-02-25", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Reasoning preview built on Qwen2.5-Max", "source": "https://qwenlm.github.io/blog/qwq-max-preview/" }, { "name": "Wan 2.1", "org": "Alibaba", "country": "China", "date": "2025-02-25", "precision": "day", "category": "video", "open_weights": true, "params": "1.3B, 14B", "note": "Open video generation suite topping VBench", "source": "https://github.com/Wan-Video/Wan2.1" }, { "name": "YandexGPT 5", "org": "Yandex", "country": "Russia", "date": "2025-02-25", "precision": "day", "category": "language", "open_weights": true, "params": "8B (Lite)", "note": "Pro model plus open-weight 8B Lite pretrain", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Granite 3.2", "org": "IBM", "country": "USA", "date": "2025-02-26", "precision": "day", "category": "language", "open_weights": true, "params": "2B, 8B", "note": "Open models adding toggleable chain-of-thought reasoning", "source": "https://www.ibm.com/new/announcements/ibm-granite-3-2-open-source-reasoning-and-vision" }, { "name": "Granite Vision 3.2", "org": "IBM", "country": "USA", "date": "2025-02-26", "precision": "day", "category": "language", "open_weights": true, "params": "2B", "note": "Document-understanding vision-language model", "source": "https://www.ibm.com/new/announcements/ibm-granite-3-2-open-source-reasoning-and-vision" }, { "name": "Hi Robot", "org": "Physical Intelligence", "country": "USA", "date": "2025-02-26", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Hierarchical VLA following open-ended instructions and feedback", "source": "https://arxiv.org/abs/2502.19417" }, { "name": "MegaTTS 3", "org": "ByteDance", "country": "China", "date": "2025-02-26", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Lightweight zero-shot voice-cloning TTS; code/weights opened March 2025", "source": "https://arxiv.org/abs/2502.18924" }, { "name": "Mercury Coder", "org": "Inception Labs", "country": "USA", "date": "2025-02-26", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "First commercial-scale diffusion LLM, over 1,000 tokens/sec", "source": "https://techcrunch.com/2025/02/26/inception-emerges-from-stealth-with-a-new-type-of-ai-model/" }, { "name": "Octave", "org": "Hume AI", "country": "USA", "date": "2025-02-26", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "LLM-based TTS that understands what it's saying", "source": "https://www.hume.ai/blog/octave-the-first-text-to-speech-model-that-understands-what-its-saying" }, { "name": "Phi-4-mini", "org": "Microsoft", "country": "USA", "date": "2025-02-26", "precision": "day", "category": "language", "open_weights": true, "params": "3.8B", "note": "Compact model with strong reasoning and function calling", "source": "https://azure.microsoft.com/en-us/blog/empowering-innovation-the-next-generation-of-the-phi-family/" }, { "name": "Phi-4-multimodal", "org": "Microsoft", "country": "USA", "date": "2025-02-26", "precision": "day", "category": "multimodal", "open_weights": true, "params": "5.6B", "note": "Speech, vision and text in one small model", "source": "https://azure.microsoft.com/en-us/blog/empowering-innovation-the-next-generation-of-the-phi-family/" }, { "name": "Scribe", "org": "ElevenLabs", "country": "USA", "date": "2025-02-26", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "ElevenLabs' first speech-to-text model, 99 languages", "source": "https://elevenlabs.io/blog/meet-scribe" }, { "name": "Command R7B Arabic", "org": "Cohere", "country": "Canada", "date": "2025-02-27", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Compact Arabic-optimized enterprise model", "source": "https://huggingface.co/CohereLabs/c4ai-command-r7b-arabic-02-2025" }, { "name": "Conversational Speech Model (CSM)", "org": "Sesame", "country": "USA", "date": "2025-02-27", "precision": "day", "category": "audio", "open_weights": true, "params": "1B (open)", "note": "Viral lifelike voice demo; 1B model open-sourced March 2025", "source": "https://www.sesame.com/research/crossing_the_uncanny_valley_of_voice" }, { "name": "GPT-4.5", "org": "OpenAI", "country": "USA", "date": "2025-02-27", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "OpenAI's largest chat model; research preview", "source": "https://en.wikipedia.org/wiki/GPT-4.5" }, { "name": "Hunyuan Turbo S", "org": "Tencent", "country": "China", "date": "2025-02-27", "precision": "day", "category": "language", "open_weights": false, "params": "560B (MoE)", "note": "Hybrid Mamba-Transformer fast-thinking model with near-instant replies", "source": "https://winbuzzer.com/2025/02/27/tencent-unveils-hunyuan-turbo-s-model-to-beat-deepseek-r1-with-near-iinstant-replies-xcxwbn/" }, { "name": "Kanana", "org": "Kakao", "country": "South Korea", "date": "2025-02-27", "precision": "day", "category": "language", "open_weights": true, "params": "2.1B (Nano)", "note": "Kakao's model family; Nano 2.1B open-sourced with tech report", "source": "https://huggingface.co/kakaocorp/kanana-nano-2.1b-instruct" }, { "name": "Pika 2.2", "org": "Pika", "country": "USA", "date": "2025-02-27", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Pikaframes keyframe transitions and 10-second 1080p clips", "source": "https://www.linkedin.com/posts/pika-labs_launch-alert-pika-22-is-here-with-pikaframeskey-activity-7300923033542688770-BRfc" }, { "name": "RoboBrain", "org": "BAAI", "country": "China", "date": "2025-02-28", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Unified brain model for robotic manipulation", "source": "https://arxiv.org/abs/2502.21257" }, { "name": "Reve Image 1.0", "org": "Reve", "country": "USA", "date": "2025-03-01", "precision": "month", "category": "image", "open_weights": false, "params": "", "note": "Debuted as 'Halfmoon', topping image-generation arena leaderboards", "source": "https://en.wikipedia.org/wiki/Text-to-image_model" }, { "name": "Aya Vision", "org": "Cohere", "country": "Canada", "date": "2025-03-04", "precision": "day", "category": "language", "open_weights": true, "params": "8B, 32B", "note": "Multilingual vision-language model covering 23 languages", "source": "https://cohere.com/blog/aya-vision" }, { "name": "CogView4", "org": "Zhipu AI", "country": "China", "date": "2025-03-04", "precision": "day", "category": "image", "open_weights": true, "params": "6B", "note": "Open bilingual text-to-image model with Chinese text rendering", "source": "https://github.com/zai-org/CogView4" }, { "name": "Instella", "org": "AMD", "country": "USA", "date": "2025-03-05", "precision": "day", "category": "language", "open_weights": true, "params": "3B", "note": "Fully open 3B models trained on MI300X GPUs", "source": "https://rocm.blogs.amd.com/artificial-intelligence/introducing-instella-3B/README.html" }, { "name": "Character-3", "org": "Hedra", "country": "USA", "date": "2025-03-06", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Omnimodal model animating expressive talking characters", "source": "https://www.linkedin.com/posts/hedra-labs_introducing-hedra-studio-and-character-3-activity-7303470158843379712-Gjxy" }, { "name": "HunyuanVideo-I2V", "org": "Tencent", "country": "China", "date": "2025-03-06", "precision": "day", "category": "video", "open_weights": true, "params": "13B", "note": "Open image-to-video extension of HunyuanVideo", "source": "https://github.com/Tencent-Hunyuan/HunyuanVideo-I2V" }, { "name": "Jamba 1.6", "org": "AI21 Labs", "country": "Israel", "date": "2025-03-06", "precision": "day", "category": "language", "open_weights": true, "params": "52B (Mini) / 398B (Large)", "note": "Hybrid SSM-Transformer open models for enterprise", "source": "https://www.ai21.com/blog/introducing-jamba-1-6/" }, { "name": "Mistral OCR", "org": "Mistral AI", "country": "France", "date": "2025-03-06", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Document understanding API for PDFs and images", "source": "https://mistral.ai/news/mistral-ocr" }, { "name": "QwQ-32B", "org": "Alibaba", "country": "China", "date": "2025-03-06", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32B", "note": "RL-trained 32B reasoning model rivaling DeepSeek-R1", "source": "https://qwenlm.github.io/blog/qwq-32b/" }, { "name": "Gemini Embedding", "org": "Google DeepMind", "country": "UK", "date": "2025-03-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Gemini-based text embedding topping MTEB multilingual", "source": "https://developers.googleblog.com/en/gemini-embedding-text-model-now-available-gemini-api/" }, { "name": "Ling-Lite / Ling-Plus", "org": "Ant Group", "country": "China", "date": "2025-03-07", "precision": "day", "category": "language", "open_weights": true, "params": "16.8B / 290B", "note": "MoE models trained without premium GPUs", "source": "https://arxiv.org/abs/2503.05139" }, { "name": "FoxBrain", "org": "Foxconn", "country": "Taiwan", "date": "2025-03-10", "precision": "day", "category": "language", "open_weights": true, "params": "70B", "note": "Traditional-Chinese reasoning LLM built on Llama 3.1", "source": "https://huggingface.co/FoxconnAI/Llama_3.1-FoxBrain-70B" }, { "name": "Genie Operator-1 (GO-1)", "org": "AgiBot", "country": "China", "date": "2025-03-10", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Generalist embodied foundation model using ViLLA latent-action framework", "source": "https://www.globenewswire.com/news-release/2025/03/10/3040128/0/en/agibot-go-1-the-evolution-of-generalist-embodied-foundation-model-from-vla-to-villa.html" }, { "name": "Reka Flash 3", "org": "Reka AI", "country": "USA", "date": "2025-03-10", "precision": "day", "category": "reasoning", "open_weights": true, "params": "21B", "note": "Open 21B general-purpose reasoning model", "source": "https://www.reka.ai/news/introducing-reka-flash" }, { "name": "Seedream 2.0", "org": "ByteDance", "country": "China", "date": "2025-03-10", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Bilingual image model with native Chinese text rendering", "source": "https://arxiv.org/abs/2503.07703" }, { "name": "GPT-4o Search Preview", "org": "OpenAI", "country": "USA", "date": "2025-03-11", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Search-tuned GPT-4o and GPT-4o mini API models launched with Responses API", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "OlympicCoder", "org": "Hugging Face", "country": "USA", "date": "2025-03-11", "precision": "day", "category": "code", "open_weights": true, "params": "7B, 32B", "note": "Open reasoning models for olympiad programming", "source": "https://huggingface.co/blog/open-r1/update-3" }, { "name": "Gemini 2.0 Flash native image generation", "org": "Google DeepMind", "country": "UK", "date": "2025-03-12", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Experimental conversational image generation and editing in Gemini", "source": "https://developers.googleblog.com/en/experiment-with-gemini-20-flash-native-image-generation/" }, { "name": "Gemini Robotics", "org": "Google DeepMind", "country": "UK", "date": "2025-03-12", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Gemini 2.0-based vision-language-action model for robots", "source": "https://deepmind.google/discover/blog/gemini-robotics-brings-ai-into-the-physical-world/" }, { "name": "Gemini Robotics-ER", "org": "Google DeepMind", "country": "UK", "date": "2025-03-12", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Embodied-reasoning model for spatial understanding", "source": "https://deepmind.google/discover/blog/gemini-robotics-brings-ai-into-the-physical-world/" }, { "name": "Gemma 3", "org": "Google DeepMind", "country": "UK", "date": "2025-03-12", "precision": "day", "category": "language", "open_weights": true, "params": "1B-27B", "note": "Open multimodal family; 27B led open models on Chatbot Arena", "source": "https://blog.google/technology/developers/gemma-3/" }, { "name": "Marey", "org": "Moonvalley", "country": "USA", "date": "2025-03-12", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Video model trained only on owned or licensed footage", "source": "https://techcrunch.com/2025/03/12/moonvalley-releases-a-video-generator-it-claims-was-trained-on-licensed-content/" }, { "name": "Open-Sora 2.0", "org": "HPC-AI Tech", "country": "Singapore", "date": "2025-03-12", "precision": "day", "category": "video", "open_weights": true, "params": "11B", "note": "Open video model trained for about $200K", "source": "https://github.com/hpcaitech/Open-Sora" }, { "name": "ShieldGemma 2", "org": "Google DeepMind", "country": "UK", "date": "2025-03-12", "precision": "day", "category": "vision", "open_weights": true, "params": "4B", "note": "Image safety classifier built on Gemma 3", "source": "https://blog.google/technology/developers/gemma-3/" }, { "name": "The AI Scientist-v2", "org": "Sakana AI", "country": "Japan", "date": "2025-03-12", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Produced first fully AI-generated paper to pass peer review", "source": "https://sakana.ai/ai-scientist-first-publication/" }, { "name": "Command A", "org": "Cohere", "country": "Canada", "date": "2025-03-13", "precision": "day", "category": "language", "open_weights": true, "params": "111B", "note": "Enterprise model running on just two GPUs", "source": "https://cohere.com/blog/command-a" }, { "name": "OLMo 2 32B", "org": "Allen Institute for AI", "country": "USA", "date": "2025-03-13", "precision": "day", "category": "language", "open_weights": true, "params": "32B", "note": "First fully open model to beat GPT-3.5 Turbo and GPT-4o mini", "source": "https://allenai.org/blog/olmo2-32B" }, { "name": "VGGT", "org": "Meta", "country": "USA", "date": "2025-03-14", "precision": "day", "category": "vision", "open_weights": true, "params": "1B", "note": "Feed-forward 3D reconstruction; CVPR 2025 best paper", "source": "https://arxiv.org/abs/2503.11651" }, { "name": "ERNIE 4.5", "org": "Baidu", "country": "China", "date": "2025-03-16", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Native multimodal foundation model; weights later opened June 30, 2025", "source": "https://techcrunch.com/2025/03/16/baidu-launches-two-new-versions-of-its-ai-model-ernie/" }, { "name": "ERNIE X1", "org": "Baidu", "country": "China", "date": "2025-03-16", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Reasoning model priced at half of DeepSeek-R1", "source": "https://techcrunch.com/2025/03/16/baidu-launches-two-new-versions-of-its-ai-model-ernie/" }, { "name": "Mistral Small 3.1", "org": "Mistral AI", "country": "France", "date": "2025-03-17", "precision": "day", "category": "language", "open_weights": true, "params": "24B", "note": "Adds vision and 128K context to Small 3", "source": "https://mistral.ai/news/mistral-small-3-1" }, { "name": "Step-Video-TI2V", "org": "StepFun", "country": "China", "date": "2025-03-17", "precision": "day", "category": "video", "open_weights": true, "params": "30B", "note": "Open text-driven image-to-video model", "source": "https://github.com/stepfun-ai/Step-Video-TI2V" }, { "name": "Cosmos-Reason1", "org": "NVIDIA", "country": "USA", "date": "2025-03-18", "precision": "day", "category": "reasoning", "open_weights": true, "params": "7B, 56B", "note": "Physical common-sense and embodied reasoning models", "source": "https://arxiv.org/abs/2503.15558" }, { "name": "Cosmos-Transfer1", "org": "NVIDIA", "country": "USA", "date": "2025-03-18", "precision": "day", "category": "video", "open_weights": true, "params": "7B", "note": "Multi-control conditional world generation for physical AI", "source": "https://arxiv.org/abs/2503.14492" }, { "name": "EXAONE Deep", "org": "LG AI Research", "country": "South Korea", "date": "2025-03-18", "precision": "day", "category": "reasoning", "open_weights": true, "params": "2.4B, 7.8B, 32B", "note": "LG's reasoning-enhanced open models", "source": "https://the-decoder.com/exaone-deep-lg-ai-research-releases-reasoning-models/" }, { "name": "Isaac GR00T N1", "org": "NVIDIA", "country": "USA", "date": "2025-03-18", "precision": "day", "category": "robotics", "open_weights": true, "params": "2B", "note": "First open humanoid robot foundation model", "source": "https://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks" }, { "name": "Llama Nemotron Nano", "org": "NVIDIA", "country": "USA", "date": "2025-03-18", "precision": "day", "category": "reasoning", "open_weights": true, "params": "8B", "note": "Llama-derived model with toggleable reasoning", "source": "https://nvidianews.nvidia.com/news/nvidia-launches-family-of-open-reasoning-ai-models-for-developers-and-enterprises-to-build-agentic-ai-platforms" }, { "name": "Llama Nemotron Super", "org": "NVIDIA", "country": "USA", "date": "2025-03-18", "precision": "day", "category": "reasoning", "open_weights": true, "params": "49B", "note": "Pruned Llama 3.3 reasoning model for a single GPU", "source": "https://nvidianews.nvidia.com/news/nvidia-launches-family-of-open-reasoning-ai-models-for-developers-and-enterprises-to-build-agentic-ai-platforms" }, { "name": "Llama Nemotron Ultra", "org": "NVIDIA", "country": "USA", "date": "2025-03-18", "precision": "day", "category": "reasoning", "open_weights": true, "params": "253B", "note": "Announced at GTC; weights released April 7, rivaling DeepSeek-R1", "source": "https://huggingface.co/nvidia/Llama-3_1-Nemotron-Ultra-253B-v1" }, { "name": "Orpheus TTS", "org": "Canopy Labs", "country": "USA", "date": "2025-03-18", "precision": "day", "category": "audio", "open_weights": true, "params": "3B", "note": "Open Llama-based TTS with human-like emotion", "source": "https://huggingface.co/canopylabs/orpheus-3b-0.1-ft" }, { "name": "Skywork R1V", "org": "Kunlun Tech", "country": "China", "date": "2025-03-18", "precision": "day", "category": "reasoning", "open_weights": true, "params": "38B", "note": "Open multimodal reasoning model", "source": "https://huggingface.co/Skywork/Skywork-R1V-38B" }, { "name": "Stable Virtual Camera", "org": "Stability AI", "country": "UK", "date": "2025-03-18", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Turns images into 3D camera-controlled multi-view video", "source": "https://stability.ai/news/introducing-stable-virtual-camera-multi-view-video-generation-with-3d-camera-control" }, { "name": "Udio v1.5 Allegro", "org": "Udio", "country": "USA", "date": "2025-03-18", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Faster, improved Udio music generation model", "source": "https://en.wikipedia.org/wiki/Udio" }, { "name": "gpt-4o-mini-transcribe", "org": "OpenAI", "country": "USA", "date": "2025-03-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Cheaper GPT-4o mini speech-to-text variant", "source": "https://techcrunch.com/2025/03/20/openai-upgrades-its-transcription-and-voice-generating-ai-models/" }, { "name": "gpt-4o-mini-tts", "org": "OpenAI", "country": "USA", "date": "2025-03-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Steerable TTS: developers instruct how it speaks", "source": "https://techcrunch.com/2025/03/20/openai-upgrades-its-transcription-and-voice-generating-ai-models/" }, { "name": "gpt-4o-transcribe", "org": "OpenAI", "country": "USA", "date": "2025-03-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "GPT-4o-based speech-to-text outperforming Whisper", "source": "https://techcrunch.com/2025/03/20/openai-upgrades-its-transcription-and-voice-generating-ai-models/" }, { "name": "RF-DETR", "org": "Roboflow", "country": "USA", "date": "2025-03-20", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Real-time transformer object detector topping COCO benchmarks", "source": "https://blog.roboflow.com/rf-detr/" }, { "name": "Hunyuan-T1", "org": "Tencent", "country": "China", "date": "2025-03-21", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Official reasoning model on hybrid Mamba-Transformer base", "source": "https://www.sohu.com/a/874263101_114760" }, { "name": "MoshiVis", "org": "Kyutai", "country": "France", "date": "2025-03-21", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Real-time speech model that can talk about images", "source": "https://www.marktechpost.com/2025/03/21/kyutai-releases-moshivis-the-first-open-source-real-time-speech-model-that-can-talk-about-images/" }, { "name": "Nemotron-H", "org": "NVIDIA", "country": "USA", "date": "2025-03-21", "precision": "day", "category": "language", "open_weights": true, "params": "8B, 47B, 56B", "note": "Hybrid Mamba-Transformer model family", "source": "https://research.nvidia.com/labs/adlr/nemotronh/" }, { "name": "DeepSeek-V3-0324", "org": "DeepSeek", "country": "China", "date": "2025-03-24", "precision": "day", "category": "language", "open_weights": true, "params": "671B (37B active)", "note": "MIT-licensed V3 upgrade with much stronger reasoning and coding", "source": "https://api-docs.deepseek.com/news/news250325" }, { "name": "Qwen2.5-VL-32B", "org": "Alibaba", "country": "China", "date": "2025-03-24", "precision": "day", "category": "language", "open_weights": true, "params": "32B", "note": "RL-tuned 32B VLM surpassing its 72B sibling", "source": "https://qwenlm.github.io/blog/qwen2.5-vl-32b/" }, { "name": "Gemini 2.5 Pro", "org": "Google DeepMind", "country": "UK", "date": "2025-03-25", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Thinking model debuting first on LMArena by wide margin", "source": "https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/" }, { "name": "GPT-4o image generation", "org": "OpenAI", "country": "USA", "date": "2025-03-25", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Native image generation replacing DALL-E; sparked Ghibli-style craze", "source": "https://openai.com/index/introducing-4o-image-generation/" }, { "name": "TxGemma", "org": "Google DeepMind", "country": "UK", "date": "2025-03-25", "precision": "day", "category": "science", "open_weights": true, "params": "2B, 9B, 27B", "note": "Open Gemma models for therapeutics development", "source": "https://developers.googleblog.com/en/introducing-txgemma-open-models-improving-therapeutics-development/" }, { "name": "GAIA-2", "org": "Wayve", "country": "UK", "date": "2025-03-26", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Multi-view generative world model for driving simulation", "source": "https://wayve.ai/thinking/gaia-2/" }, { "name": "Ideogram 3.0", "org": "Ideogram", "country": "Canada", "date": "2025-03-26", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Style references and improved photorealism and text rendering", "source": "https://about.ideogram.ai/3.0" }, { "name": "Mureka O1", "org": "Kunlun Tech", "country": "China", "date": "2025-03-26", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Billed as the first music reasoning model", "source": "https://www.newswire.ca/news-releases/kunlun-tech-launches-the-world-s-first-music-reasoning-large-model-mureka-o1-leading-the-global-ai-music-revolution-810790922.html" }, { "name": "Qwen2.5-Omni", "org": "Alibaba", "country": "China", "date": "2025-03-26", "precision": "day", "category": "multimodal", "open_weights": true, "params": "7B", "note": "End-to-end Thinker-Talker omni model: sees, hears, speaks", "source": "https://arxiv.org/abs/2503.20215" }, { "name": "QVQ-Max", "org": "Alibaba", "country": "China", "date": "2025-03-28", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Visual reasoning model that thinks with evidence", "source": "https://qwenlm.github.io/blog/qvq-max-preview/" }, { "name": "Amazon Nova Act", "org": "Amazon", "country": "USA", "date": "2025-03-31", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Browser-action agent model with research-preview SDK", "source": "https://labs.amazon.science/blog/nova-act" }, { "name": "AutoGLM Rumination", "org": "Zhipu AI", "country": "China", "date": "2025-03-31", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Free deep-research agent built on GLM rumination model", "source": "https://technode.com/2025/04/01/zhipu-ai-launches-free-ai-agent-as-chinas-tech-race-heats-up/" }, { "name": "Gen-4", "org": "Runway", "country": "USA", "date": "2025-03-31", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Consistent characters and scenes across shots", "source": "https://techcrunch.com/2025/03/31/runway-releases-an-impressive-new-video-generating-ai-model/" }, { "name": "Step-R1-V-Mini", "org": "StepFun", "country": "China", "date": "2025-04-01", "precision": "month", "category": "reasoning", "open_weights": false, "params": "", "note": "Multimodal reasoning model for visual interpretation", "source": "https://en.wikipedia.org/wiki/StepFun" }, { "name": "Dream 7B", "org": "HKU, Huawei Noah's Ark Lab", "country": "China", "date": "2025-04-02", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Open diffusion language model with strong planning ability", "source": "https://hkunlp.github.io/blog/2025/dream/" }, { "name": "DreamActor-M1", "org": "ByteDance", "country": "China", "date": "2025-04-02", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Holistic human image animation with hybrid guidance", "source": "https://arxiv.org/abs/2504.01724" }, { "name": "Speech-02", "org": "MiniMax", "country": "China", "date": "2025-04-02", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Hyper-realistic multilingual TTS in turbo and HD variants", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "Midjourney V7", "org": "Midjourney", "country": "USA", "date": "2025-04-03", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "First new Midjourney model in nearly a year; adds Draft Mode", "source": "https://techcrunch.com/2025/04/03/midjourney-releases-its-first-new-ai-image-model-in-nearly-a-year/" }, { "name": "Sec-Gemini v1", "org": "Google", "country": "USA", "date": "2025-04-04", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Experimental cybersecurity-specialized Gemini model", "source": "https://security.googleblog.com/2025/04/google-launches-sec-gemini-v1-new.html" }, { "name": "WHAMM", "org": "Microsoft", "country": "USA", "date": "2025-04-04", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Real-time playable AI-generated Quake II world model", "source": "https://www.microsoft.com/en-us/research/articles/whamm-real-time-world-modelling-of-interactive-environments/" }, { "name": "Llama 4 Behemoth", "org": "Meta", "country": "USA", "date": "2025-04-05", "precision": "day", "category": "language", "open_weights": false, "params": "~2T (288B active)", "note": "Teacher model previewed while still training; never released", "source": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "name": "Llama 4 Maverick", "org": "Meta", "country": "USA", "date": "2025-04-05", "precision": "day", "category": "language", "open_weights": true, "params": "400B (17B active)", "note": "128-expert MoE flagship open model", "source": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "name": "Llama 4 Scout", "org": "Meta", "country": "USA", "date": "2025-04-05", "precision": "day", "category": "language", "open_weights": true, "params": "109B (17B active)", "note": "Natively multimodal MoE with 10M-token context", "source": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "name": "Amazon Nova Reel 1.1", "org": "Amazon", "country": "USA", "date": "2025-04-07", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Multi-shot videos up to two minutes long", "source": "https://aws.amazon.com/blogs/aws/amazon-nova-reel-1-1-featuring-up-to-2-minutes-multi-shot-videos/" }, { "name": "Gen-4 Turbo", "org": "Runway", "country": "USA", "date": "2025-04-07", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Runway's fastest, most efficient Gen-4 video variant", "source": "https://runway.com/changelog" }, { "name": "HiDream-I1", "org": "HiDream.ai", "country": "China", "date": "2025-04-07", "precision": "day", "category": "image", "open_weights": true, "params": "17B", "note": "Open 17B image model topping arena benchmarks", "source": "https://github.com/HiDream-ai/HiDream-I1" }, { "name": "Amazon Nova Sonic", "org": "Amazon", "country": "USA", "date": "2025-04-08", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Unified speech-to-speech conversational model", "source": "https://aws.amazon.com/blogs/aws/introducing-amazon-nova-sonic-human-like-voice-conversations-for-generative-ai-applications/" }, { "name": "Cogito v1 Preview", "org": "Deep Cogito", "country": "USA", "date": "2025-04-08", "precision": "day", "category": "reasoning", "open_weights": true, "params": "3B-70B", "note": "Hybrid reasoning models trained by iterated distillation", "source": "https://techcrunch.com/2025/04/08/deep-cogito-emerges-from-stealth-with-hybrid-ai-reasoning-models/" }, { "name": "DeepCoder-14B-Preview", "org": "Agentica, Together AI", "country": "USA", "date": "2025-04-08", "precision": "day", "category": "code", "open_weights": true, "params": "14B", "note": "Fully open RL-trained coder matching o3-mini", "source": "https://www.together.ai/blog/deepcoder" }, { "name": "Kimi-VL", "org": "Moonshot AI", "country": "China", "date": "2025-04-10", "precision": "day", "category": "language", "open_weights": true, "params": "16B (2.8B active)", "note": "Efficient MoE VLM with long-thinking variant", "source": "https://arxiv.org/abs/2504.07491" }, { "name": "Pangu Ultra", "org": "Huawei", "country": "China", "date": "2025-04-10", "precision": "day", "category": "language", "open_weights": false, "params": "135B", "note": "Dense 135B model trained entirely on Ascend NPUs", "source": "https://arxiv.org/abs/2504.07866" }, { "name": "Seed-Thinking-v1.5", "org": "ByteDance", "country": "China", "date": "2025-04-10", "precision": "day", "category": "reasoning", "open_weights": false, "params": "200B (20B active)", "note": "Reasoning MoE behind Doubao-1.5-thinking-pro", "source": "https://arxiv.org/abs/2504.13914" }, { "name": "SenseNova V6", "org": "SenseTime", "country": "China", "date": "2025-04-10", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Native multimodal model with long chain-of-thought reasoning", "source": "https://www.prnewswire.com/apac/news-releases/sensetimes-sensenova-v6-chinas-most-advanced-multimodal-model-with-the-lowest-cost-in-the-industry-302426998.html" }, { "name": "ZR1-1.5B", "org": "Zyphra", "country": "USA", "date": "2025-04-10", "precision": "day", "category": "reasoning", "open_weights": true, "params": "1.5B", "note": "Small RL-trained math and code reasoning model", "source": "https://www.zyphra.com/post/introducing-zr1-1-5b-a-small-but-powerful-math-code-reasoning-model" }, { "name": "InternVL3", "org": "Shanghai AI Lab", "country": "China", "date": "2025-04-11", "precision": "day", "category": "language", "open_weights": true, "params": "1B-78B", "note": "Open VLMs rivaling GPT-4o on multimodal benchmarks", "source": "https://internvl.github.io/blog/2025-04-11-InternVL-3.0/" }, { "name": "Seaweed-7B", "org": "ByteDance", "country": "China", "date": "2025-04-11", "precision": "day", "category": "video", "open_weights": false, "params": "7B", "note": "Cost-effective 7B video generation foundation model", "source": "https://arxiv.org/abs/2504.08685" }, { "name": "Skywork-OR1", "org": "Kunlun Tech", "country": "China", "date": "2025-04-13", "precision": "day", "category": "reasoning", "open_weights": true, "params": "7B, 32B", "note": "Open RL reasoning models for math and code", "source": "https://github.com/SkyworkAI/Skywork-OR1" }, { "name": "C2S-Scale", "org": "Google", "country": "USA", "date": "2025-04-14", "precision": "day", "category": "science", "open_weights": true, "params": "up to 27B", "note": "Cell2Sentence LLMs for single-cell biology, with Yale", "source": "https://www.biorxiv.org/content/10.1101/2025.04.14.648850v2" }, { "name": "DolphinGemma", "org": "Google", "country": "USA", "date": "2025-04-14", "precision": "day", "category": "audio", "open_weights": false, "params": "~400M", "note": "Models dolphin vocalizations to help decode communication", "source": "https://blog.google/technology/ai/dolphingemma/" }, { "name": "GLM-4-32B-0414", "org": "Zhipu AI", "country": "China", "date": "2025-04-14", "precision": "day", "category": "language", "open_weights": true, "params": "9B, 32B", "note": "MIT-licensed GLM-4 models rivaling GPT-4o", "source": "https://github.com/zai-org/GLM-4" }, { "name": "GLM-Z1-32B-0414", "org": "Zhipu AI", "country": "China", "date": "2025-04-14", "precision": "day", "category": "reasoning", "open_weights": true, "params": "9B, 32B", "note": "Open deep-thinking models plus Z1-Rumination research variant", "source": "https://huggingface.co/zai-org/GLM-Z1-32B-0414" }, { "name": "GPT-4.1", "org": "OpenAI", "country": "USA", "date": "2025-04-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "1M-token API model focused on coding and instruction following", "source": "https://en.wikipedia.org/wiki/GPT-4.1" }, { "name": "GPT-4.1 mini", "org": "OpenAI", "country": "USA", "date": "2025-04-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Matches GPT-4o on many tasks at far lower cost", "source": "https://techcrunch.com/2025/04/14/openais-new-gpt-4-1-models-focus-on-coding/" }, { "name": "GPT-4.1 nano", "org": "OpenAI", "country": "USA", "date": "2025-04-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "OpenAI's fastest, cheapest model at launch", "source": "https://techcrunch.com/2025/04/14/openais-new-gpt-4-1-models-focus-on-coding/" }, { "name": "Embed 4", "org": "Cohere", "country": "Canada", "date": "2025-04-15", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Multimodal enterprise embedding model with 128K context", "source": "https://cohere.com/blog/embed-4" }, { "name": "Kimina-Prover Preview", "org": "Moonshot AI", "country": "China", "date": "2025-04-15", "precision": "day", "category": "science", "open_weights": true, "params": "72B", "note": "RL-trained Lean 4 theorem prover, built with Numina", "source": "https://arxiv.org/abs/2504.11354" }, { "name": "Kling 2.0 Master", "org": "Kuaishou", "country": "China", "date": "2025-04-15", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Kling 2.0 with multimodal visual-language editing", "source": "https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-advances-20-era-empowering-everyone-tell-great-stories" }, { "name": "Kolors 2.0", "org": "Kuaishou", "country": "China", "date": "2025-04-15", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Image model upgrade launched alongside Kling 2.0", "source": "https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-advances-20-era-empowering-everyone-tell-great-stories" }, { "name": "Seedream 3.0", "org": "ByteDance", "country": "China", "date": "2025-04-15", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Bilingual image model with native 2K output", "source": "https://arxiv.org/abs/2504.11346" }, { "name": "TerraMind", "org": "IBM", "country": "USA", "date": "2025-04-15", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Any-to-any generative Earth observation model, built with ESA", "source": "https://arxiv.org/abs/2504.11171" }, { "name": "BitNet b1.58 2B4T", "org": "Microsoft", "country": "USA", "date": "2025-04-16", "precision": "day", "category": "language", "open_weights": true, "params": "2B", "note": "First open native 1-bit LLM at 2B scale", "source": "https://arxiv.org/abs/2504.12285" }, { "name": "Granite 3.3", "org": "IBM", "country": "USA", "date": "2025-04-16", "precision": "day", "category": "language", "open_weights": true, "params": "2B, 8B", "note": "Refined reasoning, RAG LoRAs and fill-in-the-middle", "source": "https://www.ibm.com/new/announcements/ibm-granite-3-3-speech-recognition-refined-reasoning-rag-loras" }, { "name": "Granite Speech 3.3", "org": "IBM", "country": "USA", "date": "2025-04-16", "precision": "day", "category": "audio", "open_weights": true, "params": "8B", "note": "IBM's first official Granite speech-to-text model", "source": "https://www.ibm.com/new/announcements/ibm-granite-3-3-speech-recognition-refined-reasoning-rag-loras" }, { "name": "MAI-DS-R1", "org": "Microsoft", "country": "USA", "date": "2025-04-16", "precision": "day", "category": "reasoning", "open_weights": true, "params": "671B (37B active)", "note": "DeepSeek-R1 post-trained for safety and fewer refusals", "source": "https://huggingface.co/microsoft/MAI-DS-R1" }, { "name": "o4-mini", "org": "OpenAI", "country": "USA", "date": "2025-04-16", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Fast, cost-efficient reasoning with top AIME scores", "source": "https://en.wikipedia.org/wiki/OpenAI_o4-mini" }, { "name": "ProGen3", "org": "Profluent", "country": "USA", "date": "2025-04-16", "precision": "day", "category": "science", "open_weights": false, "params": "339M-46B", "note": "Protein language model family showing scaling laws for protein design", "source": "https://www.profluent.bio/showcase/progen3" }, { "name": "UI-TARS-1.5", "org": "ByteDance", "country": "China", "date": "2025-04-16", "precision": "day", "category": "agent", "open_weights": true, "params": "7B (open)", "note": "Open GUI agent model leading OSWorld and game tasks", "source": "https://seed.bytedance.com/blog/bytedance-seed-agent-model-ui-tars-1-5-open-source-achieving-sota-performance-in-various-benchmarks" }, { "name": "Gemini 2.5 Flash", "org": "Google DeepMind", "country": "UK", "date": "2025-04-17", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "First hybrid-reasoning Gemini with adjustable thinking budget", "source": "https://developers.googleblog.com/en/start-building-with-gemini-25-flash/" }, { "name": "Meta Locate 3D", "org": "Meta", "country": "USA", "date": "2025-04-17", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "End-to-end 3D object localization from natural language", "source": "https://ai.meta.com/blog/meta-fair-updates-perception-localization-reasoning/" }, { "name": "Perception Encoder", "org": "Meta", "country": "USA", "date": "2025-04-17", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Large-scale open vision encoder for images and video", "source": "https://ai.meta.com/blog/meta-fair-updates-perception-localization-reasoning/" }, { "name": "Perception Language Model (PLM)", "org": "Meta", "country": "USA", "date": "2025-04-17", "precision": "day", "category": "language", "open_weights": true, "params": "1B, 3B, 8B", "note": "Fully reproducible open vision-language model family", "source": "https://ai.meta.com/blog/meta-fair-updates-perception-localization-reasoning/" }, { "name": "Wan2.1-FLF2V", "org": "Alibaba", "country": "China", "date": "2025-04-17", "precision": "day", "category": "video", "open_weights": true, "params": "14B", "note": "Open first-and-last-frame-to-video model", "source": "https://huggingface.co/Wan-AI/Wan2.1-FLF2V-14B-720P" }, { "name": "Gemma 3 QAT", "org": "Google DeepMind", "country": "UK", "date": "2025-04-18", "precision": "day", "category": "language", "open_weights": true, "params": "1B-27B", "note": "Quantization-aware Gemma 3 runs 27B on consumer GPUs", "source": "https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/" }, { "name": "Dia", "org": "Nari Labs", "country": "South Korea", "date": "2025-04-21", "precision": "day", "category": "audio", "open_weights": true, "params": "1.6B", "note": "Two-person startup's open TTS for realistic multi-speaker dialogue", "source": "https://huggingface.co/nari-labs/Dia-1.6B" }, { "name": "MAGI-1", "org": "Sand AI", "country": "China", "date": "2025-04-21", "precision": "day", "category": "video", "open_weights": true, "params": "24B", "note": "Open autoregressive video generation at scale", "source": "https://github.com/SandAI-org/MAGI-1" }, { "name": "SkyReels-V2", "org": "Kunlun Tech", "country": "China", "date": "2025-04-21", "precision": "day", "category": "video", "open_weights": true, "params": "1.3B-14B", "note": "Infinite-length film generation via diffusion forcing", "source": "https://github.com/SkyworkAI/SkyReels-V2" }, { "name": "Vidu Q1", "org": "ShengShu", "country": "China", "date": "2025-04-21", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Top-ranked image-to-video with first-last-frame control", "source": "https://www.prnewswire.com/news/shengshu-technology/" }, { "name": "π0.5", "org": "Physical Intelligence", "country": "USA", "date": "2025-04-22", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "VLA generalizing to cleaning never-seen homes", "source": "https://arxiv.org/abs/2504.16054" }, { "name": "gpt-image-1", "org": "OpenAI", "country": "USA", "date": "2025-04-23", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "ChatGPT's native image model released to the API", "source": "https://en.wikipedia.org/wiki/GPT_Image" }, { "name": "Skywork R1V2", "org": "Kunlun Tech", "country": "China", "date": "2025-04-23", "precision": "day", "category": "reasoning", "open_weights": true, "params": "38B", "note": "Multimodal reasoning trained with hybrid MPO+GRPO RL", "source": "https://arxiv.org/abs/2504.16656" }, { "name": "Firefly Image Model 4", "org": "Adobe", "country": "USA", "date": "2025-04-24", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Adobe's fourth-generation commercially safe image model", "source": "https://blog.adobe.com/en/publish/2025/04/24/adobe-firefly-next-evolution-creative-ai-is-here" }, { "name": "Firefly Image Model 4 Ultra", "org": "Adobe", "country": "USA", "date": "2025-04-24", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Highest-detail variant of Firefly Image Model 4", "source": "https://blog.adobe.com/en/publish/2025/04/24/adobe-firefly-next-evolution-creative-ai-is-here" }, { "name": "HyperCLOVA X SEED", "org": "Naver", "country": "South Korea", "date": "2025-04-24", "precision": "day", "category": "language", "open_weights": true, "params": "0.5B-3B", "note": "Naver's first open-sourced HyperCLOVA X models", "source": "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B" }, { "name": "Lyria 2", "org": "Google DeepMind", "country": "UK", "date": "2025-04-24", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "High-fidelity music generation in Music AI Sandbox", "source": "https://deepmind.google/discover/blog/music-ai-sandbox-now-with-new-features-and-broader-access/" }, { "name": "Lyria RealTime", "org": "Google DeepMind", "country": "UK", "date": "2025-04-24", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Interactive music generation steered in real time; API May 2025", "source": "https://deepmind.google/discover/blog/music-ai-sandbox-now-with-new-features-and-broader-access/" }, { "name": "Step1X-Edit", "org": "StepFun", "country": "China", "date": "2025-04-24", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open image-editing model approaching GPT-4o quality", "source": "https://github.com/stepfun-ai/Step1X-Edit" }, { "name": "ERNIE 4.5 Turbo", "org": "Baidu", "country": "China", "date": "2025-04-25", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Faster ERNIE 4.5 at roughly 80% lower price", "source": "https://the-decoder.com/baidus-new-ernie-models-are-taking-on-deepseek-and-openai-with-ultra-low-pricing/" }, { "name": "ERNIE X1 Turbo", "org": "Baidu", "country": "China", "date": "2025-04-25", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Cheaper, faster version of the X1 reasoning model", "source": "https://technode.com/2025/04/27/baidu-unveils-big-bets-on-ai-apps-and-affordable-models-at-wuhan-developer-conference/" }, { "name": "Kimi-Audio", "org": "Moonshot AI", "country": "China", "date": "2025-04-25", "precision": "day", "category": "audio", "open_weights": true, "params": "7B", "note": "Open universal audio understanding and generation model", "source": "https://github.com/MoonshotAI/Kimi-Audio" }, { "name": "Palmyra X5", "org": "Writer", "country": "USA", "date": "2025-04-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Enterprise LLM with 1M-token context", "source": "https://writer.com/engineering/long-context-palmyra-x5/" }, { "name": "Llama Guard 4", "org": "Meta", "country": "USA", "date": "2025-04-29", "precision": "day", "category": "language", "open_weights": true, "params": "12B", "note": "Multimodal safety classifier released at LlamaCon with Prompt Guard 2", "source": "https://ai.meta.com/blog/ai-defenders-program-llama-protection-tools/" }, { "name": "Qwen3", "org": "Alibaba", "country": "China", "date": "2025-04-29", "precision": "day", "category": "reasoning", "open_weights": true, "params": "0.6B-235B (22B active)", "note": "Hybrid-thinking open family led by Qwen3-235B-A22B", "source": "https://qwenlm.github.io/blog/qwen3/" }, { "name": "Amazon Nova Premier", "org": "Amazon", "country": "USA", "date": "2025-04-30", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Most capable Nova 1 model and distillation teacher", "source": "https://aws.amazon.com/blogs/aws/amazon-nova-premier-our-most-capable-model-for-complex-tasks-and-teacher-for-model-distillation/" }, { "name": "DeepSeek-Prover-V2", "org": "DeepSeek", "country": "China", "date": "2025-04-30", "precision": "day", "category": "science", "open_weights": true, "params": "7B, 671B", "note": "Lean 4 theorem prover using subgoal decomposition", "source": "https://arxiv.org/abs/2504.21801" }, { "name": "Gen-4 References", "org": "Runway", "country": "USA", "date": "2025-04-30", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Gen-4 image generation keeping characters and locations consistent", "source": "https://runway.com/changelog" }, { "name": "Mellum", "org": "JetBrains", "country": "Czech Republic", "date": "2025-04-30", "precision": "day", "category": "code", "open_weights": true, "params": "4B", "note": "JetBrains' code-completion model released openly", "source": "https://huggingface.co/JetBrains/Mellum-4b-base" }, { "name": "MiMo-7B", "org": "Xiaomi", "country": "China", "date": "2025-04-30", "precision": "day", "category": "reasoning", "open_weights": true, "params": "7B", "note": "Xiaomi's first open reasoning LLM for math and code", "source": "https://siliconangle.com/2025/04/30/china-ai-rising-xiaomi-releases-new-mimo-7b-models-deepseek-upgrades-prover-math-ai/" }, { "name": "Phi-4-mini-reasoning", "org": "Microsoft", "country": "USA", "date": "2025-04-30", "precision": "day", "category": "reasoning", "open_weights": true, "params": "3.8B", "note": "Small math-reasoning model for edge devices", "source": "https://azure.microsoft.com/en-us/blog/one-year-of-phi-small-language-models-making-big-leaps-in-ai/" }, { "name": "Phi-4-reasoning", "org": "Microsoft", "country": "USA", "date": "2025-04-30", "precision": "day", "category": "reasoning", "open_weights": true, "params": "14B", "note": "First Phi reasoning model, rivaling much larger models", "source": "https://arxiv.org/abs/2504.21318" }, { "name": "Phi-4-reasoning-plus", "org": "Microsoft", "country": "USA", "date": "2025-04-30", "precision": "day", "category": "reasoning", "open_weights": true, "params": "14B", "note": "RL-enhanced variant of Phi-4-reasoning", "source": "https://arxiv.org/abs/2504.21318" }, { "name": "Chatterbox", "org": "Resemble AI", "country": "USA", "date": "2025-05-01", "precision": "month", "category": "audio", "open_weights": true, "params": "0.5B", "note": "Open MIT-licensed TTS with emotion-exaggeration control", "source": "https://huggingface.co/ResembleAI/chatterbox" }, { "name": "Parakeet TDT 0.6B v2", "org": "NVIDIA", "country": "USA", "date": "2025-05-01", "precision": "day", "category": "audio", "open_weights": true, "params": "600M", "note": "Open ASR model topping Hugging Face Open ASR leaderboard", "source": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2" }, { "name": "PixVerse V4.5", "org": "PixVerse", "country": "China", "date": "2025-05-01", "precision": "month", "category": "video", "open_weights": false, "params": "", "note": "Alibaba-backed video generator update", "source": "https://www.veed.io/ai-models/video/pixverse-v4.5" }, { "name": "Suno v4.5", "org": "Suno", "country": "USA", "date": "2025-05-01", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Richer genres and vocals; songs up to eight minutes", "source": "https://suno.com/blog/introducing-v4-5" }, { "name": "Granite 4.0 Tiny Preview", "org": "IBM", "country": "USA", "date": "2025-05-02", "precision": "day", "category": "language", "open_weights": true, "params": "7B (1B active)", "note": "Hybrid Mamba-2/Transformer MoE preview", "source": "https://www.ibm.com/new/announcements/ibm-granite-4-0-tiny-preview-sneak-peek" }, { "name": "Ming-Lite-Omni", "org": "Ant Group", "country": "China", "date": "2025-05-04", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Open omni model unifying perception and generation; v1 final May 28", "source": "https://github.com/inclusionAI/Ming" }, { "name": "LTXV-13B", "org": "Lightricks", "country": "Israel", "date": "2025-05-05", "precision": "day", "category": "video", "open_weights": true, "params": "13B", "note": "Larger open LTX-Video model with multiscale rendering", "source": "https://github.com/Lightricks/LTX-Video" }, { "name": "ACE-Step", "org": "ACE Studio & StepFun", "country": "China", "date": "2025-05-06", "precision": "day", "category": "audio", "open_weights": true, "params": "3.5B", "note": "Open music-generation foundation model", "source": "https://github.com/ace-step/ACE-Step" }, { "name": "Apriel-Nemotron-15B-Thinker", "org": "ServiceNow", "country": "USA", "date": "2025-05-06", "precision": "day", "category": "reasoning", "open_weights": true, "params": "15B", "note": "Enterprise reasoning model built with NVIDIA at half peers' memory", "source": "https://www.marktechpost.com/2025/05/09/servicenow-ai-released-apriel-nemotron-15b-thinker-a-compact-yet-powerful-reasoning-model-optimized-for-enterprise-scale-deployment-and-efficiency/" }, { "name": "Bielik v3", "org": "SpeakLeash", "country": "Poland", "date": "2025-05-06", "precision": "day", "category": "language", "open_weights": true, "params": "1.5B / 4.5B", "note": "Small efficient Polish-language models", "source": "https://pl.wikipedia.org/wiki/Bielik_(model_językowy)" }, { "name": "Gemini 2.5 Pro Preview (I/O edition)", "org": "Google DeepMind", "country": "UK", "date": "2025-05-06", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Coding-focused upgrade topping WebDev Arena", "source": "https://blog.google/products/gemini/gemini-2-5-pro-updates" }, { "name": "Mistral Medium 3", "org": "Mistral AI", "country": "France", "date": "2025-05-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Frontier-class performance at roughly 8x lower cost", "source": "https://mistral.ai/news/mistral-medium-3" }, { "name": "Pangu Ultra MoE", "org": "Huawei", "country": "China", "date": "2025-05-07", "precision": "day", "category": "language", "open_weights": false, "params": "718B", "note": "718B sparse MoE trained entirely on Ascend NPUs", "source": "https://arxiv.org/abs/2505.04519" }, { "name": "HunyuanCustom", "org": "Tencent", "country": "China", "date": "2025-05-08", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Multimodal-driven customized video generation", "source": "https://github.com/Tencent-Hunyuan/HunyuanCustom" }, { "name": "Seed-Coder", "org": "ByteDance", "country": "China", "date": "2025-05-08", "precision": "day", "category": "code", "open_weights": true, "params": "8B", "note": "Open code LLM family with model-curated training data", "source": "https://github.com/ByteDance-Seed/Seed-Coder" }, { "name": "INTELLECT-2", "org": "Prime Intellect", "country": "USA", "date": "2025-05-11", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32B", "note": "First 32B model trained via globally distributed RL", "source": "https://www.primeintellect.ai/blog/intellect-2-release" }, { "name": "Seed1.5-VL", "org": "ByteDance", "country": "China", "date": "2025-05-11", "precision": "day", "category": "language", "open_weights": false, "params": "20B active", "note": "VLM with state-of-the-art results on 38 of 60 benchmarks", "source": "https://arxiv.org/abs/2505.07062" }, { "name": "Continuous Thought Machine (CTM)", "org": "Sakana AI", "country": "Japan", "date": "2025-05-12", "precision": "day", "category": "reasoning", "open_weights": true, "params": "", "note": "Architecture using neuron timing and synchronization to think", "source": "https://sakana.ai/ctm/" }, { "name": "Seed1.5-Embedding", "org": "ByteDance", "country": "China", "date": "2025-05-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Retrieval embedding model with SOTA Chinese and English results", "source": "https://seed.bytedance.com/blog/bytedance-s-seed1-5-embedding-model-achieves-sota-in-retrieval-training-details-unveiled" }, { "name": "Step1X-3D", "org": "StepFun", "country": "China", "date": "2025-05-13", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open textured 3D asset generator released with 800K-asset dataset", "source": "https://github.com/stepfun-ai/Step1X-3D" }, { "name": "AlphaEvolve", "org": "Google DeepMind", "country": "UK", "date": "2025-05-14", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Gemini-powered agent discovering new algorithms and math results", "source": "https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/" }, { "name": "BLIP3-o", "org": "Salesforce", "country": "USA", "date": "2025-05-14", "precision": "day", "category": "image", "open_weights": true, "params": "4B, 8B", "note": "Fully open unified image understanding and generation", "source": "https://arxiv.org/abs/2505.09568" }, { "name": "Stable Audio Open Small", "org": "Stability AI", "country": "UK", "date": "2025-05-14", "precision": "day", "category": "audio", "open_weights": true, "params": "341M", "note": "On-device text-to-audio developed with Arm", "source": "https://stability.ai/news/stability-ai-and-arm-release-stable-audio-open-small-enabling-real-world-deployment-for-on-device-audio-control" }, { "name": "UMA (Universal Model for Atoms)", "org": "Meta", "country": "USA", "date": "2025-05-14", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Universal interatomic potential released with OMol25 dataset", "source": "https://ai.meta.com/blog/meta-fair-science-new-open-source-releases/" }, { "name": "Wan2.1-VACE", "org": "Alibaba", "country": "China", "date": "2025-05-14", "precision": "day", "category": "video", "open_weights": true, "params": "1.3B, 14B", "note": "All-in-one open video creation and editing model", "source": "https://github.com/Wan-Video/Wan2.1" }, { "name": "Falcon-Edge", "org": "TII", "country": "UAE", "date": "2025-05-15", "precision": "day", "category": "language", "open_weights": true, "params": "1B / 3B", "note": "BitNet 1.58-bit ternary models for edge devices", "source": "https://falcon-lm.github.io/blog/falcon-edge/" }, { "name": "SWE-1", "org": "Windsurf", "country": "USA", "date": "2025-05-15", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Windsurf's first in-house software-engineering model family", "source": "https://windsurf.com/blog/windsurf-wave-9-swe-1" }, { "name": "codex-1", "org": "OpenAI", "country": "USA", "date": "2025-05-16", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "o3-derived model powering the cloud Codex agent", "source": "https://techcrunch.com/2025/05/16/openai-launches-codex-an-ai-coding-agent-in-chatgpt/" }, { "name": "codex-mini-latest", "org": "OpenAI", "country": "USA", "date": "2025-05-16", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "o4-mini-based low-latency coding model for Codex CLI", "source": "https://the-decoder.com/openai-launches-codex-autonomous-ai-agents-for-software-development/" }, { "name": "HunyuanImage 2.0", "org": "Tencent", "country": "China", "date": "2025-05-16", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Real-time (millisecond) text-to-image generation", "source": "https://finance.sina.com.cn/tech/discovery/2025-05-16/doc-inewttnz8576026.shtml" }, { "name": "Cosmos-Predict2", "org": "NVIDIA", "country": "USA", "date": "2025-05-18", "precision": "day", "category": "video", "open_weights": true, "params": "2B, 14B", "note": "Next-generation world foundation models for physical AI", "source": "https://nvidianews.nvidia.com/news/nvidia-powers-humanoid-robot-industry-with-cloud-to-robot-computing-platforms-for-physical-ai" }, { "name": "Isaac GR00T N1.5", "org": "NVIDIA", "country": "USA", "date": "2025-05-18", "precision": "day", "category": "robotics", "open_weights": true, "params": "3B", "note": "Humanoid foundation model update trained in 36 hours on synthetic data", "source": "https://nvidianews.nvidia.com/news/nvidia-powers-humanoid-robot-industry-with-cloud-to-robot-computing-platforms-for-physical-ai" }, { "name": "BAGEL", "org": "ByteDance", "country": "China", "date": "2025-05-20", "precision": "day", "category": "multimodal", "open_weights": true, "params": "14B (7B active)", "note": "Open unified multimodal understanding and generation model", "source": "https://arxiv.org/abs/2505.14683" }, { "name": "Falcon-H1", "org": "TII", "country": "UAE", "date": "2025-05-20", "precision": "day", "category": "language", "open_weights": true, "params": "0.5B-34B", "note": "Hybrid Transformer-Mamba open model family", "source": "https://falcon-lm.github.io/blog/falcon-h1/" }, { "name": "Gemini 2.5 Deep Think", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Parallel-thinking mode for 2.5 Pro announced at I/O; shipped August", "source": "https://techcrunch.com/2025/05/20/deep-think-boosts-the-performance-of-googles-flagship-google-gemini-ai-model/" }, { "name": "Gemini 2.5 Flash / Pro TTS", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Controllable multi-speaker native text-to-speech previews", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "Gemini 2.5 Flash Native Audio", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Native audio dialogue with emotion-aware responses", "source": "https://blog.google/technology/google-deepmind/gemini-2-5-native-audio/" }, { "name": "Gemini 2.5 Flash Preview 05-20", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "I/O update to 2.5 Flash with stronger coding and reasoning", "source": "https://blog.google/technology/ai/google-io-2025-all-our-announcements/" }, { "name": "Gemini Diffusion", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Experimental text diffusion model with very fast generation", "source": "https://deepmind.google/models/gemini-diffusion/" }, { "name": "Gemma 3n", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "multimodal", "open_weights": true, "params": "E2B, E4B (5B/8B raw)", "note": "Mobile-first audio-vision-text model; full release June 26", "source": "https://developers.googleblog.com/en/introducing-gemma-3n/" }, { "name": "Imagen 4", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Sharper detail and typography; up to 2K resolution", "source": "https://blog.google/technology/ai/generative-media-models-io-2025" }, { "name": "Imagen 4 Ultra", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Highest-fidelity Imagen 4 tier", "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/models/imagen/4-0-ultra-generate-001" }, { "name": "MedGemma", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "science", "open_weights": true, "params": "4B, 27B", "note": "Open Gemma models for medical text and image comprehension", "source": "https://ai.google.dev/gemma/docs/releases" }, { "name": "Robin", "org": "FutureHouse", "country": "USA", "date": "2025-05-20", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Multi-agent system that proposed a dry AMD drug candidate", "source": "https://www.futurehouse.org/research-announcements/demonstrating-end-to-end-scientific-discovery-with-robin-a-multi-agent-system" }, { "name": "SignGemma", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Sign-language-to-text model announced at I/O as upcoming open model", "source": "https://blog.google/technology/ai/google-io-2025-all-our-announcements/" }, { "name": "Veo 3", "org": "Google DeepMind", "country": "UK", "date": "2025-05-20", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "First widely available video model with native synchronized audio", "source": "https://blog.google/technology/ai/generative-media-models-io-2025" }, { "name": "voyage-3.5", "org": "Voyage AI", "country": "USA", "date": "2025-05-20", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Improved retrieval embeddings at lower cost", "source": "https://blog.voyageai.com/2025/05/20/voyage-3-5/" }, { "name": "Devstral", "org": "Mistral AI", "country": "France", "date": "2025-05-21", "precision": "day", "category": "code", "open_weights": true, "params": "24B", "note": "Open agentic coding model, built with All Hands AI", "source": "https://mistral.ai/news/devstral" }, { "name": "Falcon Arabic", "org": "TII", "country": "UAE", "date": "2025-05-21", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "First Arabic-focused Falcon model", "source": "https://huggingface.co/blog/tiiuae/falcon-arabic" }, { "name": "Claude Opus 4", "org": "Anthropic", "country": "USA", "date": "2025-05-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Hybrid reasoning flagship billed as world's best coding model", "source": "https://www.anthropic.com/news/claude-4" }, { "name": "Claude Sonnet 4", "org": "Anthropic", "country": "USA", "date": "2025-05-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Upgraded Sonnet with hybrid reasoning and strong coding", "source": "https://www.anthropic.com/news/claude-4" }, { "name": "Kanana 1.5", "org": "Kakao", "country": "South Korea", "date": "2025-05-23", "precision": "day", "category": "language", "open_weights": true, "params": "2.1B / 8B", "note": "Upgraded open Kakao models with better coding, math, function calling", "source": "https://github.com/kakao/kanana" }, { "name": "o3 Operator", "org": "OpenAI", "country": "USA", "date": "2025-05-23", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "o3-based computer-use model replacing GPT-4o in Operator", "source": "https://techcrunch.com/2025/05/23/openai-upgrades-the-ai-model-powering-its-operator-agent/" }, { "name": "Sarvam-M", "org": "Sarvam AI", "country": "India", "date": "2025-05-23", "precision": "day", "category": "reasoning", "open_weights": true, "params": "24B", "note": "Hybrid-reasoning Indic model built on Mistral Small", "source": "https://www.sarvam.ai/blogs/sarvam-m" }, { "name": "Pangu Pro MoE", "org": "Huawei", "country": "China", "date": "2025-05-27", "precision": "day", "category": "language", "open_weights": true, "params": "72B (16B active)", "note": "Mixture-of-grouped-experts model; weights open-sourced June 30", "source": "https://arxiv.org/abs/2505.21411" }, { "name": "PLaMo Translate", "org": "Preferred Networks", "country": "Japan", "date": "2025-05-27", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Translation-specialized model built on PLaMo 2", "source": "https://huggingface.co/api/models?author=pfnet&sort=createdAt&direction=-1&limit=60" }, { "name": "Codestral Embed", "org": "Mistral AI", "country": "France", "date": "2025-05-28", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Code-specialized embedding model for retrieval", "source": "https://mistral.ai/news/codestral-embed" }, { "name": "DeepSeek-R1-0528", "org": "DeepSeek", "country": "China", "date": "2025-05-28", "precision": "day", "category": "reasoning", "open_weights": true, "params": "671B (37B active)", "note": "Major R1 update narrowing gap with o3 and Gemini 2.5 Pro", "source": "https://api-docs.deepseek.com/news/news250528" }, { "name": "DeepSeek-R1-0528-Qwen3-8B", "org": "DeepSeek", "country": "China", "date": "2025-05-28", "precision": "day", "category": "reasoning", "open_weights": true, "params": "8B", "note": "R1-0528 reasoning distilled into Qwen3-8B", "source": "https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B" }, { "name": "HunyuanVideo-Avatar", "org": "Tencent", "country": "China", "date": "2025-05-28", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Audio-driven multi-character avatar animation", "source": "https://github.com/Tencent-Hunyuan/HunyuanVideo-Avatar" }, { "name": "Odyssey interactive video (research preview)", "org": "Odyssey", "country": "USA", "date": "2025-05-28", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Real-time interactive world model generating a frame every 40ms", "source": "https://the-decoder.com/generative-ai-startup-odyssey-demos-interactive-ai-generated-video/" }, { "name": "EVI 3", "org": "Hume AI", "country": "USA", "date": "2025-05-29", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-language model generating any voice and personality", "source": "https://www.hume.ai/blog/introducing-evi-3" }, { "name": "FLUX.1 Kontext", "org": "Black Forest Labs", "country": "Germany", "date": "2025-05-29", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "In-context image generation and editing ([pro] and [max])", "source": "https://bfl.ai/announcements/flux-1-kontext" }, { "name": "Kling 2.1", "org": "Kuaishou", "country": "China", "date": "2025-05-29", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Cheaper, faster 2.x model in Standard/Pro/Master tiers", "source": "https://www.ctol.digital/news/kling-2-1-launch-affordable-fast-ai-video-generation/" }, { "name": "Darwin Gödel Machine", "org": "Sakana AI", "country": "Japan", "date": "2025-05-30", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Self-improving coding agent that rewrites its own code", "source": "https://sakana.ai/dgm/" }, { "name": "MiMo-VL-7B", "org": "Xiaomi", "country": "China", "date": "2025-05-30", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Compact VLM with strong multimodal reasoning", "source": "https://www.marktechpost.com/2025/06/02/mimo-vl-7b-a-powerful-vision-language-model-to-enhance-general-visual-understanding-and-multimodal-reasoning/" }, { "name": "Eleven v3 (alpha)", "org": "ElevenLabs", "country": "USA", "date": "2025-06-03", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Most expressive ElevenLabs TTS with inline audio tags", "source": "https://elevenlabs.io/blog/eleven-v3" }, { "name": "Holo1", "org": "H Company", "country": "France", "date": "2025-06-03", "precision": "day", "category": "agent", "open_weights": true, "params": "3B, 7B", "note": "Open web-navigation action models powering Runner H", "source": "https://huggingface.co/Hcompany/Holo1-7B" }, { "name": "Llama Nemotron Nano VL", "org": "NVIDIA", "country": "USA", "date": "2025-06-03", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Compact document-understanding VLM topping OCRBench v2", "source": "https://developer.nvidia.com/blog/new-nvidia-llama-nemotron-nano-vision-language-model-tops-ocr-benchmark-for-accuracy/" }, { "name": "OpenAudio S1", "org": "Fish Audio", "country": "USA", "date": "2025-06-03", "precision": "day", "category": "audio", "open_weights": false, "params": "4B", "note": "Emotion-controllable TTS; 0.5B distilled S1-mini variant", "source": "https://openaudio.com/blogs/s1" }, { "name": "SmolVLA", "org": "Hugging Face", "country": "USA", "date": "2025-06-03", "precision": "day", "category": "robotics", "open_weights": true, "params": "450M", "note": "Compact open VLA trained on community LeRobot data", "source": "https://huggingface.co/blog/smolvla" }, { "name": "Claude Gov", "org": "Anthropic", "country": "USA", "date": "2025-06-05", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Custom models for U.S. national security customers", "source": "https://the-decoder.com/anthropic-launches-claude-gov-an-ai-model-designed-specifically-for-u-s-national-security-agencies/" }, { "name": "Comma v0.1", "org": "EleutherAI", "country": "USA", "date": "2025-06-05", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "LLM trained only on openly licensed Common Pile text", "source": "https://blog.eleuther.ai/common-pile/" }, { "name": "ether0", "org": "FutureHouse", "country": "USA", "date": "2025-06-05", "precision": "day", "category": "science", "open_weights": true, "params": "24B", "note": "Open chemistry reasoning model trained with RL", "source": "https://www.futurehouse.org/research-announcements/ether0-a-scientific-reasoning-model-for-chemistry" }, { "name": "Gemini 2.5 Pro Preview 06-05", "org": "Google DeepMind", "country": "UK", "date": "2025-06-05", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Upgraded preview with 24-point LMArena jump; became stable GA", "source": "https://blog.google/products/gemini/gemini-2-5-pro-latest-preview/" }, { "name": "Qwen3 Embedding", "org": "Alibaba", "country": "China", "date": "2025-06-05", "precision": "day", "category": "language", "open_weights": true, "params": "0.6B-8B", "note": "Embedding and reranker series topping MTEB multilingual leaderboard", "source": "https://qwenlm.github.io/blog/qwen3-embedding/" }, { "name": "SeedEdit 3.0", "org": "ByteDance", "country": "China", "date": "2025-06-05", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Faster, more consistent image editing", "source": "https://arxiv.org/abs/2506.05083" }, { "name": "Boltz-2", "org": "MIT & Recursion", "country": "USA", "date": "2025-06-06", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open model jointly predicting structure and binding affinity", "source": "https://boltz.com/boltz2" }, { "name": "dots.llm1", "org": "Xiaohongshu", "country": "China", "date": "2025-06-06", "precision": "day", "category": "language", "open_weights": true, "params": "142B (14B active)", "note": "Rednote's first open-source MoE LLM", "source": "https://huggingface.co/rednote-hilab/dots.llm1.inst" }, { "name": "MiniCPM4", "org": "OpenBMB", "country": "China", "date": "2025-06-06", "precision": "day", "category": "language", "open_weights": true, "params": "0.5B, 8B", "note": "Edge model with 5x+ generation speedup via sparse attention", "source": "https://github.com/OpenBMB/MiniCPM" }, { "name": "RoboBrain 2.0", "org": "BAAI", "country": "China", "date": "2025-06-06", "precision": "day", "category": "robotics", "open_weights": true, "params": "32B", "note": "Open embodied brain model for humanoid robots", "source": "https://github.com/FlagOpen/RoboBrain2.0" }, { "name": "Sarvam-Translate", "org": "Sarvam AI", "country": "India", "date": "2025-06-07", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Open translation model for Indian languages", "source": "https://www.sarvam.ai/blogs/sarvam-translate" }, { "name": "Apple Foundation Models (2025)", "org": "Apple", "country": "USA", "date": "2025-06-09", "precision": "day", "category": "language", "open_weights": false, "params": "~3B on-device + server MoE", "note": "On-device model opened to developers via Foundation Models framework", "source": "https://machinelearning.apple.com/research/apple-foundation-models-2025-updates" }, { "name": "Magistral Medium", "org": "Mistral AI", "country": "France", "date": "2025-06-10", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Mistral's first reasoning model, multilingual chain of thought", "source": "https://mistral.ai/news/magistral" }, { "name": "Magistral Small", "org": "Mistral AI", "country": "France", "date": "2025-06-10", "precision": "day", "category": "reasoning", "open_weights": true, "params": "24B", "note": "Open-weight 24B reasoning model", "source": "https://mistral.ai/news/magistral" }, { "name": "o3-pro", "org": "OpenAI", "country": "USA", "date": "2025-06-10", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Higher-compute o3 for the most reliable answers", "source": "https://techcrunch.com/2025/06/10/openai-releases-o3-pro-a-souped-up-version-of-its-o3-ai-reasoning-model/" }, { "name": "Redwood AI", "org": "1X", "country": "USA", "date": "2025-06-10", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Mobile-manipulation VLA for the NEO humanoid", "source": "https://www.1x.tech/discover/redwood-ai" }, { "name": "Veo 3 Fast", "org": "Google DeepMind", "country": "UK", "date": "2025-06-10", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Faster, cheaper Veo 3 variant at 720p", "source": "https://the-decoder.com/googles-veo-3-fast-now-generates-720p-ai-videos-at-more-than-double-the-previous-speed/" }, { "name": "Seed1.6", "org": "ByteDance", "country": "China", "date": "2025-06-11", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Doubao 1.6 multimodal deep-thinking family at one-third prior cost", "source": "https://www.aibase.com/news/18831" }, { "name": "Seedance 1.0", "org": "ByteDance", "country": "China", "date": "2025-06-11", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Multi-shot video model rivaling Veo 3 on leaderboards", "source": "https://seed.bytedance.com/blog/tech-report-of-seedance-1-0-is-now-publicly-available" }, { "name": "V-JEPA 2", "org": "Meta", "country": "USA", "date": "2025-06-11", "precision": "day", "category": "vision", "open_weights": true, "params": "1.2B", "note": "Self-supervised video world model enabling zero-shot robot planning", "source": "https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/" }, { "name": "Weather Lab cyclone model (FGN)", "org": "Google DeepMind", "country": "UK", "date": "2025-06-12", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Predicts cyclone formation, track, intensity 15 days ahead; used by NHC", "source": "https://deepmind.google/discover/blog/weather-lab-cyclone-predictions-with-ai/" }, { "name": "Hunyuan3D 2.1", "org": "Tencent", "country": "China", "date": "2025-06-13", "precision": "day", "category": "image", "open_weights": true, "params": "3.3B", "note": "Production-ready open 3D model with PBR materials", "source": "https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1" }, { "name": "Kimi-Dev-72B", "org": "Moonshot AI", "country": "China", "date": "2025-06-16", "precision": "day", "category": "code", "open_weights": true, "params": "72B", "note": "60.4% SWE-bench Verified, top open model at release", "source": "https://moonshotai.github.io/Kimi-Dev/" }, { "name": "MiniMax-M1", "org": "MiniMax", "country": "China", "date": "2025-06-16", "precision": "day", "category": "reasoning", "open_weights": true, "params": "456B (45.9B active)", "note": "Open hybrid-attention reasoning model with 1M context", "source": "https://www.minimax.io/news/minimaxm1" }, { "name": "OmniGen2", "org": "BAAI", "country": "China", "date": "2025-06-16", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open unified image generation and editing model", "source": "https://github.com/VectorSpaceLab/OmniGen2" }, { "name": "ALE-Agent", "org": "Sakana AI", "country": "Japan", "date": "2025-06-17", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Optimization agent placing 21st of 1,000+ humans in AtCoder contest", "source": "https://sakana.ai/ale-bench/" }, { "name": "Gemini 2.5 Flash-Lite", "org": "Google DeepMind", "country": "UK", "date": "2025-06-17", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Fastest, cheapest 2.5 model; Pro and Flash reach GA same day", "source": "https://blog.google/products/gemini/gemini-2-5-model-family-expands/" }, { "name": "AFM-4.5B", "org": "Arcee AI", "country": "USA", "date": "2025-06-18", "precision": "day", "category": "language", "open_weights": true, "params": "4.5B", "note": "Arcee's first in-house pretrained foundation model", "source": "https://www.arcee.ai/blog/announcing-the-arcee-foundation-model-family" }, { "name": "Hailuo 02", "org": "MiniMax", "country": "China", "date": "2025-06-18", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Native 1080p video model with strong physics at low cost", "source": "https://www.minimax.io/news/minimax-hailuo-02" }, { "name": "Midjourney V1 Video", "org": "Midjourney", "country": "USA", "date": "2025-06-18", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Midjourney's first image-to-video model", "source": "https://updates.midjourney.com/introducing-our-v1-video-model/" }, { "name": "Skala", "org": "Microsoft", "country": "USA", "date": "2025-06-18", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Deep-learning exchange-correlation functional for accurate DFT", "source": "https://www.microsoft.com/en-us/research/blog/breaking-bonds-breaking-ground-advancing-the-accuracy-of-computational-chemistry-with-deep-learning/" }, { "name": "Kyutai STT", "org": "Kyutai", "country": "France", "date": "2025-06-19", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open streaming speech-to-text models powering Unmute", "source": "https://kyutai.org/blog" }, { "name": "Hunyuan-GameCraft", "org": "Tencent", "country": "China", "date": "2025-06-20", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Interactive game-video world model; paper June 2025, weights August 2025", "source": "https://arxiv.org/abs/2506.17201" }, { "name": "Kimi-Researcher", "org": "Moonshot AI", "country": "China", "date": "2025-06-20", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "End-to-end RL-trained autonomous research agent", "source": "https://moonshotai.github.io/Kimi-Researcher/" }, { "name": "Magenta RealTime", "org": "Google DeepMind", "country": "UK", "date": "2025-06-20", "precision": "day", "category": "audio", "open_weights": true, "params": "800M", "note": "Open-weights live music generation model", "source": "https://magenta.withgoogle.com/magenta-realtime" }, { "name": "Mistral Small 3.2", "org": "Mistral AI", "country": "France", "date": "2025-06-20", "precision": "day", "category": "language", "open_weights": true, "params": "24B", "note": "Better instruction following and function calling", "source": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506" }, { "name": "Music-1.5", "org": "MiniMax", "country": "China", "date": "2025-06-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music generation from prompts and lyrics; full release adds 4-minute songs", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "Pangu 5.5", "org": "Huawei", "country": "China", "date": "2025-06-20", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "HDC 2025 upgrade with reasoning Pangu Ultra MoE", "source": "https://en.wikipedia.org/wiki/Huawei_PanGu" }, { "name": "Mu", "org": "Microsoft", "country": "USA", "date": "2025-06-23", "precision": "day", "category": "language", "open_weights": false, "params": "330M", "note": "On-device model powering the Windows Settings agent", "source": "https://blogs.windows.com/windowsexperience/2025/06/23/introducing-mu-language-model-and-how-it-enabled-the-agent-in-windows-settings/" }, { "name": "State", "org": "Arc Institute", "country": "USA", "date": "2025-06-23", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Arc's first virtual cell model predicting perturbation responses", "source": "https://arcinstitute.org/news/virtual-cell-model-state" }, { "name": "Gemini Robotics On-Device", "org": "Google DeepMind", "country": "UK", "date": "2025-06-24", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "VLA model running locally on robots", "source": "https://deepmind.google/discover/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/" }, { "name": "o3-deep-research", "org": "OpenAI", "country": "USA", "date": "2025-06-24", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Deep research model exposed in the API", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "o4-mini-deep-research", "org": "OpenAI", "country": "USA", "date": "2025-06-24", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Cheaper, faster deep research model in the API", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "AlphaGenome", "org": "Google DeepMind", "country": "UK", "date": "2025-06-25", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Predicts regulatory effects of DNA variants over 1Mb sequences", "source": "https://deepmind.google/discover/blog/alphagenome-ai-for-better-understanding-the-genome/" }, { "name": "DiffuCoder", "org": "Apple", "country": "USA", "date": "2025-06-25", "precision": "day", "category": "code", "open_weights": true, "params": "7B", "note": "Masked diffusion model for code generation", "source": "https://arxiv.org/abs/2506.20639" }, { "name": "jina-embeddings-v4", "org": "Jina AI", "country": "Germany", "date": "2025-06-25", "precision": "day", "category": "language", "open_weights": true, "params": "3.8B", "note": "Universal multimodal, multilingual retrieval embeddings", "source": "https://jina.ai/news/jina-embeddings-v4-universal-embeddings-for-multimodal-multilingual-retrieval/" }, { "name": "FLUX.1 Kontext [dev]", "org": "Black Forest Labs", "country": "Germany", "date": "2025-06-26", "precision": "day", "category": "image", "open_weights": true, "params": "12B", "note": "Open-weight in-context image editing model", "source": "https://bfl.ai/announcements/flux-1-kontext-dev" }, { "name": "Kwai Keye-VL", "org": "Kuaishou", "country": "China", "date": "2025-06-26", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Open multimodal LLM specialized for short-video understanding", "source": "https://github.com/Kwai-Keye/Keye" }, { "name": "Qwen VLo", "org": "Alibaba", "country": "China", "date": "2025-06-26", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Unified multimodal understanding and image generation", "source": "https://qwenlm.github.io/blog/qwen-vlo/" }, { "name": "Hunyuan-A13B", "org": "Tencent", "country": "China", "date": "2025-06-27", "precision": "day", "category": "reasoning", "open_weights": true, "params": "80B (13B active)", "note": "Open MoE with fast/slow thinking and 256K context", "source": "https://github.com/Tencent-Hunyuan/Hunyuan-A13B" }, { "name": "HyperCLOVA X THINK", "org": "Naver", "country": "South Korea", "date": "2025-06-27", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Naver's reasoning-focused model described in technical report", "source": "https://arxiv.org/abs/2506.22403" }, { "name": "Qwen-TTS", "org": "Alibaba", "country": "China", "date": "2025-06-27", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "TTS supporting Pekingese, Shanghainese and Sichuanese dialects", "source": "https://qwenlm.github.io/blog/qwen-tts/" }, { "name": "Chai-2", "org": "Chai Discovery", "country": "USA", "date": "2025-06-30", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Zero-shot antibody design with roughly 16% hit rate", "source": "https://www.chaidiscovery.com/news/introducing-chai-2" }, { "name": "ERNIE 4.5 (open-source family)", "org": "Baidu", "country": "China", "date": "2025-06-30", "precision": "day", "category": "language", "open_weights": true, "params": "0.3B-424B", "note": "Baidu open-sources ten ERNIE 4.5 models incl. multimodal MoE", "source": "https://technode.com/2025/07/01/baidu-open-sources-ernie-4-5-series-models-including-multimodal-moe-architecture/" }, { "name": "MAI-DxO", "org": "Microsoft", "country": "USA", "date": "2025-06-30", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Diagnostic orchestrator solving 85% of NEJM case challenges", "source": "https://microsoft.ai/new/the-path-to-medical-superintelligence/" }, { "name": "openPangu Embedded 7B", "org": "Huawei", "country": "China", "date": "2025-06-30", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Huawei's first open-weight Pangu release, alongside Pro MoE weights", "source": "http://www.news.cn/tech/20250630/ce015889d3b9400aa179179ac8a5aeb0/c.html" }, { "name": "Act-Two", "org": "Runway", "country": "USA", "date": "2025-07-01", "precision": "month", "category": "video", "open_weights": false, "params": "", "note": "Motion-capture model driving characters from head, face, body and hand performance", "source": "https://runway.com/changelog" }, { "name": "GLM-4.1V-Thinking", "org": "Zhipu AI", "country": "China", "date": "2025-07-01", "precision": "day", "category": "reasoning", "open_weights": true, "params": "9B", "note": "Open 9B vision-language reasoning model trained with RL curriculum sampling", "source": "https://github.com/zai-org/GLM-V" }, { "name": "Goedel-Prover-V2", "org": "Princeton University", "country": "USA", "date": "2025-07-01", "precision": "month", "category": "science", "open_weights": true, "params": "32B", "note": "Open Lean prover with scaffolded data synthesis and self-correction", "source": "https://arxiv.org/abs/2508.03613" }, { "name": "HyperCLOVA X SEED Think 14B", "org": "NAVER", "country": "South Korea", "date": "2025-07-01", "precision": "month", "category": "reasoning", "open_weights": true, "params": "14B", "note": "NAVER's open Korean hybrid-reasoning model", "source": "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Think-14B" }, { "name": "MuseSteamer", "org": "Baidu", "country": "China", "date": "2025-07-02", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Baidu's first video generator, with native Chinese audio", "source": "https://www.reuters.com/technology/baidu-launches-ai-video-generator-overhauls-search-features-2025-07-02/" }, { "name": "A.X 4.0", "org": "SK Telecom", "country": "South Korea", "date": "2025-07-03", "precision": "day", "category": "language", "open_weights": true, "params": "72B / 7B", "note": "Korean-optimized enterprise models built on Qwen2.5", "source": "https://huggingface.co/skt/A.X-4.0" }, { "name": "Jamba 1.7", "org": "AI21 Labs", "country": "Israel", "date": "2025-07-03", "precision": "day", "category": "language", "open_weights": true, "params": "398B (94B active)", "note": "Hybrid SSM-Transformer update with improved grounding and instruction following", "source": "https://docs.ai21.com/changelog" }, { "name": "Kyutai TTS", "org": "Kyutai", "country": "France", "date": "2025-07-03", "precision": "day", "category": "audio", "open_weights": true, "params": "1.6B", "note": "Streaming text-to-speech, open-sourced with the Unmute voice stack", "source": "https://kyutai.org/blog" }, { "name": "Mi:dm 2.0", "org": "KT", "country": "South Korea", "date": "2025-07-04", "precision": "day", "category": "language", "open_weights": true, "params": "11.5B", "note": "KT's Korea-centric open models in Base 11.5B and Mini sizes", "source": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct" }, { "name": "SmolLM3", "org": "Hugging Face", "country": "USA", "date": "2025-07-08", "precision": "day", "category": "language", "open_weights": true, "params": "3B", "note": "Fully open 3B multilingual long-context model with dual thinking modes", "source": "https://huggingface.co/blog/smollm3" }, { "name": "FlexOlmo", "org": "Allen Institute for AI", "country": "USA", "date": "2025-07-09", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Training paradigm letting data owners collaborate without sharing data", "source": "https://allenai.org/blog/flexolmo" }, { "name": "Grok 4", "org": "xAI", "country": "USA", "date": "2025-07-09", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "xAI flagship reasoning model launched alongside multi-agent Heavy tier", "source": "https://x.ai/news/grok-4" }, { "name": "Grok 4 Heavy", "org": "xAI", "country": "USA", "date": "2025-07-09", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Parallel test-time-compute Grok 4; first to 50% on Humanity's Last Exam", "source": "https://x.ai/news/grok-4" }, { "name": "MedGemma 27B Multimodal", "org": "Google", "country": "USA", "date": "2025-07-09", "precision": "day", "category": "language", "open_weights": true, "params": "27B", "note": "Open multimodal medical model handling health records and medical images", "source": "https://research.google/blog/medgemma-our-most-capable-open-models-for-health-ai-development/" }, { "name": "MedSigLIP", "org": "Google", "country": "USA", "date": "2025-07-09", "precision": "day", "category": "vision", "open_weights": true, "params": "0.8B", "note": "Medical image-text encoder for classification and retrieval", "source": "https://research.google/blog/medgemma-our-most-capable-open-models-for-health-ai-development/" }, { "name": "Phi-4-mini-flash-reasoning", "org": "Microsoft", "country": "USA", "date": "2025-07-09", "precision": "day", "category": "reasoning", "open_weights": true, "params": "3.8B", "note": "Hybrid SambaY architecture for fast on-device math reasoning", "source": "https://azure.microsoft.com/en-us/blog/reasoning-reimagined-introducing-phi-4-mini-flash-reasoning/" }, { "name": "T5Gemma", "org": "Google DeepMind", "country": "USA", "date": "2025-07-09", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Collection of encoder-decoder models adapted from Gemma 2", "source": "https://developers.googleblog.com/en/t5gemma/" }, { "name": "Audio Flamingo 3", "org": "NVIDIA", "country": "USA", "date": "2025-07-10", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open audio-language model for speech, sound and music", "source": "https://huggingface.co/nvidia/audio-flamingo-3-hf" }, { "name": "Devstral Medium", "org": "Mistral AI", "country": "France", "date": "2025-07-10", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "API agentic coding model released alongside Devstral Small 1.1", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "Devstral Small 1.1", "org": "Mistral AI", "country": "France", "date": "2025-07-10", "precision": "day", "category": "code", "open_weights": true, "params": "24B", "note": "Open agentic software-engineering model update", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "LFM2", "org": "Liquid AI", "country": "USA", "date": "2025-07-10", "precision": "day", "category": "language", "open_weights": true, "params": "1.2B", "note": "Liquid AI's second-generation fast on-device foundation models", "source": "https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models" }, { "name": "Reka Flash 3.1", "org": "Reka", "country": "USA", "date": "2025-07-10", "precision": "day", "category": "reasoning", "open_weights": true, "params": "21B", "note": "Improved open reasoning model released with Reka Quant quantization", "source": "https://www.reka.ai/news/reka-flash-3-1-and-reka-quant" }, { "name": "Solar Pro 2", "org": "Upstage", "country": "South Korea", "date": "2025-07-10", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Frontier Korean model with reasoning mode", "source": "https://www.upstage.ai/blog/en/solar-pro-2-launch" }, { "name": "Kimi K2", "org": "Moonshot AI", "country": "China", "date": "2025-07-11", "precision": "day", "category": "language", "open_weights": true, "params": "1T (32B active)", "note": "Open trillion-parameter MoE built for agentic tool use and coding", "source": "https://moonshotai.github.io/Kimi-K2/" }, { "name": "CogVideoX-3", "org": "Zhipu AI", "country": "China", "date": "2025-07-15", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Video generation upgrade adding start- and end-frame synthesis", "source": "https://docs.z.ai/release-notes/new-released" }, { "name": "EXAONE 4.0", "org": "LG AI Research", "country": "South Korea", "date": "2025-07-15", "precision": "day", "category": "language", "open_weights": true, "params": "32B", "note": "LG's hybrid LLM unifying fast and deep-thinking modes", "source": "https://github.com/LG-AI-EXAONE/EXAONE-4.0" }, { "name": "Voxtral", "org": "Mistral AI", "country": "France", "date": "2025-07-15", "precision": "day", "category": "audio", "open_weights": true, "params": "24B", "note": "Mistral's first open speech-understanding models (Small 24B, Mini 3B)", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "LTX-Video 0.9.8", "org": "Lightricks", "country": "Israel", "date": "2025-07-16", "precision": "day", "category": "video", "open_weights": true, "params": "13B", "note": "Autoregressive update enabling open video generation up to 60 seconds", "source": "https://siliconangle.com/2025/07/16/lightricks-latest-release-allows-creators-direct-longform-ai-generated-videos-real-time/" }, { "name": "Param-1", "org": "BharatGen", "country": "India", "date": "2025-07-16", "precision": "day", "category": "language", "open_weights": true, "params": "2.9B", "note": "Government-backed bilingual Hindi-English Indian foundation model", "source": "https://arxiv.org/abs/2507.13390" }, { "name": "Canary-Qwen-2.5B", "org": "NVIDIA", "country": "USA", "date": "2025-07-17", "precision": "day", "category": "audio", "open_weights": true, "params": "2.5B", "note": "Speech-augmented LLM topping the Open ASR leaderboard", "source": "https://huggingface.co/nvidia/canary-qwen-2.5b" }, { "name": "ChatGPT agent", "org": "OpenAI", "country": "USA", "date": "2025-07-17", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Merged Operator and deep research into an agentic ChatGPT mode", "source": "https://techcrunch.com/2025/07/17/openai-launches-a-general-purpose-agent-in-chatgpt/" }, { "name": "MirageLSD", "org": "Decart", "country": "Israel", "date": "2025-07-17", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Real-time live-stream video-to-video transformation model", "source": "https://decart.ai/blog" }, { "name": "OpenReasoning-Nemotron", "org": "NVIDIA", "country": "USA", "date": "2025-07-18", "precision": "day", "category": "reasoning", "open_weights": true, "params": "1.5B-32B", "note": "Reasoning models distilled from DeepSeek-R1-0528 traces", "source": "https://huggingface.co/blog/nvidia/openreasoning-nemotron" }, { "name": "Seed-X", "org": "ByteDance", "country": "China", "date": "2025-07-18", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Open 7B multilingual translation model rivaling far larger models", "source": "https://arxiv.org/abs/2507.13618" }, { "name": "T-Pro 2.0", "org": "T-Bank", "country": "Russia", "date": "2025-07-18", "precision": "day", "category": "language", "open_weights": true, "params": "33B", "note": "Russian open hybrid-reasoning model built on Qwen3-32B", "source": "https://www.tbank.ru/about/news/18072025-t-technologies-group-has-released-model-with-hybrid-reasoning-mode-t-pro-2-0/" }, { "name": "IMO gold-medal experimental reasoning model", "org": "OpenAI", "country": "USA", "date": "2025-07-19", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Unreleased general LLM scored gold-medal level at IMO 2025", "source": "https://en.wikipedia.org/wiki/2025_in_artificial_intelligence" }, { "name": "Gemini Deep Think (IMO gold)", "org": "Google DeepMind", "country": "USA", "date": "2025-07-21", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "First officially certified IMO gold-medal performance by an AI system", "source": "https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/" }, { "name": "Qwen3-235B-A22B-Instruct-2507", "org": "Alibaba", "country": "China", "date": "2025-07-21", "precision": "day", "category": "language", "open_weights": true, "params": "235B (22B active)", "note": "Split Qwen3 into separate instruct and thinking models; 256K context", "source": "https://github.com/QwenLM/Qwen3" }, { "name": "GR-3", "org": "ByteDance", "country": "China", "date": "2025-07-22", "precision": "day", "category": "robotics", "open_weights": false, "params": "4B", "note": "Generalist VLA for long-horizon dexterous and bimanual manipulation", "source": "https://seed.bytedance.com/blog/seed-research-gr-3-released-a-generalist-robot-model-for-generalization-long-horizon-tasks-and-bi-manual-deformable-object-manipulation" }, { "name": "Higgs Audio v2", "org": "Boson AI", "country": "USA", "date": "2025-07-22", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Expressive open audio foundation model trained on 10M hours", "source": "https://www.boson.ai/blog/higgs-tts-2" }, { "name": "Latent-X", "org": "Latent Labs", "country": "UK", "date": "2025-07-22", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "De novo protein binder design model from Latent Labs", "source": "https://www.latentlabs.com/latent-x/" }, { "name": "Qwen3-Coder", "org": "Alibaba", "country": "China", "date": "2025-07-22", "precision": "day", "category": "code", "open_weights": true, "params": "480B (35B active)", "note": "Open agentic coding MoE launched alongside Qwen Code CLI", "source": "https://qwenlm.github.io/blog/qwen3-coder/" }, { "name": "Aeneas", "org": "Google DeepMind", "country": "USA", "date": "2025-07-23", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Contextualizes and restores ancient Latin inscriptions for historians", "source": "https://deepmind.google/blog/aeneas-transforms-how-historians-connect-the-past/" }, { "name": "Seed-Prover", "org": "ByteDance", "country": "China", "date": "2025-07-23", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Lean prover fully solved 4 of 6 IMO 2025 problems", "source": "https://seed.bytedance.com/en/blog/bytedance-seed-prover-achieves-silver-medal-score-in-imo-2025" }, { "name": "Step-Audio 2", "org": "StepFun", "country": "China", "date": "2025-07-23", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "End-to-end speech model with paralinguistic understanding and tool calling", "source": "https://github.com/stepfun-ai/Step-Audio2" }, { "name": "Kanana-1.5-15.7B-A3B", "org": "Kakao", "country": "South Korea", "date": "2025-07-24", "precision": "day", "category": "language", "open_weights": true, "params": "15.7B (3B active)", "note": "Kakao's first mixture-of-experts model", "source": "https://github.com/kakao/kanana" }, { "name": "Magistral 1.1", "org": "Mistral AI", "country": "France", "date": "2025-07-24", "precision": "day", "category": "reasoning", "open_weights": true, "params": "24B", "note": "Updated Magistral Small and Medium reasoning models", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "Qwen-MT", "org": "Alibaba", "country": "China", "date": "2025-07-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "RL-tuned translation model covering 92 languages", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Seed LiveInterpret 2.0", "org": "ByteDance", "country": "China", "date": "2025-07-24", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "End-to-end simultaneous speech-to-speech interpretation with voice cloning", "source": "https://seed.bytedance.com/en/seed_liveinterpret" }, { "name": "Llama Nemotron Super v1.5", "org": "NVIDIA", "country": "USA", "date": "2025-07-25", "precision": "day", "category": "reasoning", "open_weights": true, "params": "49B", "note": "Improved agentic reasoning update of Llama-3.3-based Nemotron Super", "source": "https://developer.nvidia.com/blog/build-more-accurate-and-efficient-ai-agents-with-the-new-nvidia-llama-nemotron-super-v1-5/" }, { "name": "Qwen3-235B-A22B-Thinking-2507", "org": "Alibaba", "country": "China", "date": "2025-07-25", "precision": "day", "category": "reasoning", "open_weights": true, "params": "235B (22B active)", "note": "Thinking-only update of the Qwen3 235B flagship", "source": "https://github.com/QwenLM/Qwen3" }, { "name": "Runway Aleph", "org": "Runway", "country": "USA", "date": "2025-07-25", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "In-context video model for editing and transforming existing footage", "source": "https://runway.com/research/introducing-runway-aleph" }, { "name": "Step 3", "org": "StepFun", "country": "China", "date": "2025-07-25", "precision": "day", "category": "reasoning", "open_weights": true, "params": "321B (38B active)", "note": "Open multimodal reasoning MoE optimized for cheap decoding on domestic chips", "source": "https://huggingface.co/stepfun-ai/step3" }, { "name": "HunyuanWorld 1.0", "org": "Tencent", "country": "China", "date": "2025-07-26", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "First open-source simulation-capable immersive 3D world generation model", "source": "https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0" }, { "name": "Intern-S1", "org": "Shanghai AI Lab", "country": "China", "date": "2025-07-26", "precision": "day", "category": "science", "open_weights": true, "params": "241B (28B active)", "note": "Open scientific multimodal reasoning model launched at WAIC 2025", "source": "https://hub.baai.ac.cn/view/47633" }, { "name": "SenseNova V6.5", "org": "SenseTime", "country": "China", "date": "2025-07-27", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Multimodal reasoning model unveiled at WAIC 2025", "source": "https://www.scmp.com/tech/article/3319751/waic-shanghai-tencent-sensetime-launch-new-ai-models-stir-industry-rivalry" }, { "name": "GLM-4.5", "org": "Zhipu AI", "country": "China", "date": "2025-07-28", "precision": "day", "category": "language", "open_weights": true, "params": "355B (32B active)", "note": "Open agentic hybrid-reasoning MoE; Zhipu rebranded as Z.ai", "source": "https://docs.z.ai/release-notes/new-released" }, { "name": "GLM-4.5-Air", "org": "Zhipu AI", "country": "China", "date": "2025-07-28", "precision": "day", "category": "language", "open_weights": true, "params": "106B (12B active)", "note": "Compact sibling of GLM-4.5 for cheap and local deployment", "source": "https://z.ai/blog/glm-4.5" }, { "name": "Grok Imagine", "org": "xAI", "country": "USA", "date": "2025-07-28", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Image and short video generation in Grok; wider launch August 4-5", "source": "https://en.wikipedia.org/wiki/Grok_(chatbot)" }, { "name": "Mureka V7", "org": "Skywork", "country": "China", "date": "2025-07-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music generation model launched with Mureka TTS V1", "source": "https://www.thatericalper.com/2025/07/28/skywork-unveils-mureka-v7-and-tts-v1-ushering-in-a-new-era-of-emotionally-intelligent-ai-music-and-voice/" }, { "name": "Wan2.2", "org": "Alibaba", "country": "China", "date": "2025-07-28", "precision": "day", "category": "video", "open_weights": true, "params": "27B (14B active)", "note": "Brought mixture-of-experts architecture into open video diffusion models", "source": "https://github.com/Wan-Video/Wan2.2" }, { "name": "Meta CLIP 2", "org": "Meta", "country": "USA", "date": "2025-07-29", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Recipe for training CLIP on worldwide multilingual web data", "source": "https://arxiv.org/abs/2507.22062" }, { "name": "Qwen3-30B-A3B-Instruct-2507", "org": "Alibaba", "country": "China", "date": "2025-07-29", "precision": "day", "category": "language", "open_weights": true, "params": "30B (3B active)", "note": "Small MoE refresh popular for local inference", "source": "https://simonwillison.net/2025/Jul/29/qwen3-30b-a3b-instruct-2507/" }, { "name": "Skild Brain", "org": "Skild AI", "country": "USA", "date": "2025-07-29", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "General-purpose robot foundation model across quadrupeds, humanoids and arms", "source": "https://www.skild.ai/blogs/building-the-general-purpose-robotic-brain" }, { "name": "AlphaEarth Foundations", "org": "Google DeepMind", "country": "USA", "date": "2025-07-30", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Virtual-satellite embedding model mapping Earth's land and coastal waters", "source": "https://deepmind.google/blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/" }, { "name": "Codestral 25.08", "org": "Mistral AI", "country": "France", "date": "2025-07-30", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Enterprise code-completion model update", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "Qwen3-30B-A3B-Thinking-2507", "org": "Alibaba", "country": "China", "date": "2025-07-30", "precision": "day", "category": "reasoning", "open_weights": true, "params": "30B (3B active)", "note": "Thinking companion to the small Qwen3-30B-A3B 2507 refresh", "source": "https://simonwillison.net/2025/Jul/30/qwen3-30b-a3b-thinking-2507/" }, { "name": "Cogito v2 Preview", "org": "Deep Cogito", "country": "USA", "date": "2025-07-31", "precision": "day", "category": "reasoning", "open_weights": true, "params": "671B (37B active)", "note": "Hybrid reasoning models from 70B to 671B trained by iterated distillation", "source": "https://www.deepcogito.com/research/cogito-v2-preview" }, { "name": "Command A Vision", "org": "Cohere", "country": "Canada", "date": "2025-07-31", "precision": "day", "category": "language", "open_weights": true, "params": "112B", "note": "Enterprise vision-language model for documents and charts", "source": "https://cohere.com/blog/command-a-vision" }, { "name": "FLUX.1 Krea [dev]", "org": "Black Forest Labs", "country": "Germany", "date": "2025-07-31", "precision": "day", "category": "image", "open_weights": true, "params": "12B", "note": "Opinionated photorealistic text-to-image model co-trained with Krea", "source": "https://bfl.ai/announcements/flux-1-krea-dev" }, { "name": "Qwen3-Coder-Flash", "org": "Alibaba", "country": "China", "date": "2025-07-31", "precision": "day", "category": "code", "open_weights": true, "params": "30B (3B active)", "note": "Small open agentic coding MoE for local use", "source": "https://simonwillison.net/2025/Jul/31/qwen3-coder-flash/" }, { "name": "Seed Diffusion Preview", "org": "ByteDance", "country": "China", "date": "2025-07-31", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Discrete diffusion language model generating code at 2,146 tokens/s", "source": "https://seed.bytedance.com/blog/seed-research-seed-diffusion-preview-released-a-diffusion-language-model-delivering-breakthrough-2-146-tokens-s-inference-speed" }, { "name": "ALLaM 34B", "org": "HUMAIN", "country": "Saudi Arabia", "date": "2025-08-01", "precision": "month", "category": "language", "open_weights": false, "params": "34B", "note": "Arabic-first model powering the HUMAIN Chat app", "source": "https://en.wikipedia.org/wiki/Humain" }, { "name": "Hunyuan-Large-Vision", "org": "Tencent", "country": "China", "date": "2025-08-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Top Chinese multimodal model on the LMArena vision leaderboard", "source": "https://the-decoder.com/tencents-hunyuan-large-vision-sets-a-new-benchmark-as-chinas-leading-multimodal-model/" }, { "name": "SEA-LION v4", "org": "AI Singapore", "country": "Singapore", "date": "2025-08-01", "precision": "month", "category": "language", "open_weights": true, "params": "27B", "note": "Southeast Asian multimodal language model built on Gemma 3", "source": "https://docs.sea-lion.ai/models/sea-lion-v4" }, { "name": "MiniCPM-V 4.0", "org": "ModelBest", "country": "China", "date": "2025-08-02", "precision": "day", "category": "language", "open_weights": true, "params": "4B", "note": "Edge vision-language model beating GPT-4.1-mini on image understanding", "source": "https://github.com/OpenBMB/MiniCPM-V" }, { "name": "Hunyuan 0.5B/1.8B/4B/7B", "org": "Tencent", "country": "China", "date": "2025-08-04", "precision": "day", "category": "language", "open_weights": true, "params": "0.5B-7B", "note": "Four compact open models for edge devices", "source": "https://www.artificialintelligence-news.com/news/tencent-releases-versatile-open-source-hunyuan-ai-models/" }, { "name": "MiDashengLM-7B", "org": "Xiaomi", "country": "China", "date": "2025-08-04", "precision": "day", "category": "audio", "open_weights": true, "params": "7B", "note": "Open audio-understanding LLM trained on general audio captions", "source": "https://github.com/xiaomi-research/dasheng-lm" }, { "name": "Qwen-Image", "org": "Alibaba", "country": "China", "date": "2025-08-04", "precision": "day", "category": "image", "open_weights": true, "params": "20B", "note": "Open 20B MMDiT image model excelling at complex text rendering", "source": "https://github.com/QwenLM/Qwen-Image" }, { "name": "Claude Opus 4.1", "org": "Anthropic", "country": "USA", "date": "2025-08-05", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Incremental upgrade to Claude Opus 4 for coding and agentic tasks", "source": "https://www.anthropic.com/news/claude-opus-4-1" }, { "name": "Eleven Music", "org": "ElevenLabs", "country": "USA", "date": "2025-08-05", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Text-to-music model marketed as cleared for commercial use", "source": "https://elevenlabs.io/blog/eleven-music-is-here" }, { "name": "Genie 3", "org": "Google DeepMind", "country": "USA", "date": "2025-08-05", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Real-time interactive world model generating navigable 720p environments", "source": "https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/" }, { "name": "gpt-oss-120b", "org": "OpenAI", "country": "USA", "date": "2025-08-05", "precision": "day", "category": "reasoning", "open_weights": true, "params": "117B (5.1B active)", "note": "OpenAI's first open-weight language model since GPT-2", "source": "https://openai.com/index/introducing-gpt-oss/" }, { "name": "gpt-oss-20b", "org": "OpenAI", "country": "USA", "date": "2025-08-05", "precision": "day", "category": "reasoning", "open_weights": true, "params": "21B (3.6B active)", "note": "Small open-weight OpenAI reasoner that runs in 16 GB of memory", "source": "https://openai.com/index/introducing-gpt-oss/" }, { "name": "MiniMax Speech 2.5", "org": "MiniMax", "country": "China", "date": "2025-08-06", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Multilingual TTS upgrade with broader language coverage and stronger cloning", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "Qwen3-4B-2507", "org": "Alibaba", "country": "China", "date": "2025-08-06", "precision": "day", "category": "language", "open_weights": true, "params": "4B", "note": "Final Qwen3-2507 release: small instruct and thinking models", "source": "https://github.com/QwenLM/Qwen3" }, { "name": "GPT-5", "org": "OpenAI", "country": "USA", "date": "2025-08-07", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Unified system routing between fast and reasoning models; replaced GPT-4o", "source": "https://en.wikipedia.org/wiki/GPT-5" }, { "name": "GPT-5 mini", "org": "OpenAI", "country": "USA", "date": "2025-08-07", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Cheaper, faster GPT-5 API variant", "source": "https://developers.openai.com/api/docs/models/gpt-5" }, { "name": "GPT-5 nano", "org": "OpenAI", "country": "USA", "date": "2025-08-07", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Smallest, cheapest GPT-5 API model", "source": "https://developers.openai.com/api/docs/models/gpt-5" }, { "name": "GPT-5 Pro", "org": "OpenAI", "country": "USA", "date": "2025-08-07", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Parallel test-time-compute GPT-5 variant for ChatGPT Pro users", "source": "https://en.wikipedia.org/wiki/GPT-5" }, { "name": "Perch 2.0", "org": "Google DeepMind", "country": "USA", "date": "2025-08-07", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open bioacoustics model identifying species across birds, mammals and reefs", "source": "https://deepmind.google/blog/how-ai-is-helping-advance-the-science-of-bioacoustics-to-save-endangered-species/" }, { "name": "Baichuan-M2", "org": "Baichuan", "country": "China", "date": "2025-08-11", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32B", "note": "Open medical reasoning model topping HealthBench among open models", "source": "https://huggingface.co/baichuan-inc/Baichuan-M2-32B" }, { "name": "GLM-4.5V", "org": "Zhipu AI", "country": "China", "date": "2025-08-11", "precision": "day", "category": "language", "open_weights": true, "params": "106B (12B active)", "note": "Open vision reasoning model incl. GUI-agent and video tasks", "source": "https://github.com/zai-org/GLM-V" }, { "name": "LFM2-VL", "org": "Liquid AI", "country": "USA", "date": "2025-08-12", "precision": "day", "category": "language", "open_weights": true, "params": "1.6B", "note": "Efficient on-device vision-language models", "source": "https://www.liquid.ai/blog/lfm2-vl-efficient-vision-language-models" }, { "name": "Matrix-Game 2.0", "org": "Skywork", "country": "China", "date": "2025-08-12", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open real-time interactive world model, a Genie 3 alternative", "source": "https://github.com/SkyworkAI/Matrix-Game" }, { "name": "Mistral Medium 3.1", "org": "Mistral AI", "country": "France", "date": "2025-08-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Mid-size Mistral API model update", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "MolmoAct", "org": "Allen Institute for AI", "country": "USA", "date": "2025-08-12", "precision": "day", "category": "robotics", "open_weights": true, "params": "7B", "note": "Open action reasoning model that plans robot motion in 3D", "source": "https://allenai.org/blog/molmoact" }, { "name": "OpenCUA", "org": "University of Hong Kong", "country": "China", "date": "2025-08-12", "precision": "day", "category": "agent", "open_weights": true, "params": "32B", "note": "Open computer-use agent framework, dataset and models", "source": "https://arxiv.org/abs/2508.09123" }, { "name": "Canary-1B-v2", "org": "NVIDIA", "country": "USA", "date": "2025-08-14", "precision": "day", "category": "audio", "open_weights": true, "params": "1B", "note": "Multilingual ASR and translation for 25 European languages", "source": "https://huggingface.co/nvidia/canary-1b-v2" }, { "name": "DINOv3", "org": "Meta", "country": "USA", "date": "2025-08-14", "precision": "day", "category": "vision", "open_weights": true, "params": "7B", "note": "Self-supervised vision backbone scaled to 7B parameters", "source": "https://ai.meta.com/blog/dinov3-self-supervised-vision-model/" }, { "name": "Gemma 3 270M", "org": "Google DeepMind", "country": "USA", "date": "2025-08-14", "precision": "day", "category": "language", "open_weights": true, "params": "270M", "note": "Compact Gemma built for task-specific fine-tuning", "source": "https://developers.googleblog.com/en/introducing-gemma-3-270m/" }, { "name": "Marey Realism v1.5", "org": "Moonvalley", "country": "Canada", "date": "2025-08-14", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Licensed-data 1080p filmmaking video model update", "source": "https://blog.fal.ai/moonvalleys-marey-realism-v1-5-is-now-on-fal-the-most-advanced-generative-filmmaking-yet/" }, { "name": "NextStep-1", "org": "StepFun", "country": "China", "date": "2025-08-14", "precision": "day", "category": "image", "open_weights": true, "params": "14B", "note": "Autoregressive image generation with continuous tokens", "source": "https://arxiv.org/abs/2508.10711" }, { "name": "Parakeet-TDT-0.6B-v3", "org": "NVIDIA", "country": "USA", "date": "2025-08-14", "precision": "day", "category": "audio", "open_weights": true, "params": "0.6B", "note": "Multilingual Parakeet ASR trained on the Granary dataset", "source": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3" }, { "name": "RoseTTAFold3 (RF3)", "org": "University of Washington", "country": "USA", "date": "2025-08-14", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open biomolecular structure predictor narrowing gap with AlphaFold 3", "source": "https://github.com/RosettaCommons/foundry" }, { "name": "Imagen 4 Fast", "org": "Google DeepMind", "country": "USA", "date": "2025-08-15", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Low-cost, speed-focused Imagen 4 tier", "source": "https://developers.googleblog.com/en/announcing-imagen-4-fast-and-imagen-4-family-generally-available-in-the-gemini-api/" }, { "name": "Ovis2.5", "org": "Alibaba", "country": "China", "date": "2025-08-15", "precision": "day", "category": "language", "open_weights": true, "params": "9B", "note": "Open native-resolution multimodal LLMs in 2B and 9B sizes", "source": "https://arxiv.org/abs/2508.11737" }, { "name": "NVIDIA Nemotron Nano 2", "org": "NVIDIA", "country": "USA", "date": "2025-08-18", "precision": "day", "category": "reasoning", "open_weights": true, "params": "9B", "note": "Hybrid Mamba-Transformer reasoner released with its pretraining dataset", "source": "https://research.nvidia.com/labs/adlr/NVIDIA-Nemotron-Nano-2/" }, { "name": "Qwen-Image-Edit", "org": "Alibaba", "country": "China", "date": "2025-08-18", "precision": "day", "category": "image", "open_weights": true, "params": "20B", "note": "Extends Qwen-Image text rendering to precise image editing", "source": "https://huggingface.co/Qwen/Qwen-Image-Edit" }, { "name": "FlowER", "org": "MIT", "country": "USA", "date": "2025-08-20", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Electron flow-matching model for reaction mechanism prediction, published in Nature", "source": "https://github.com/FongMunHong/FlowER" }, { "name": "Seed-OSS-36B", "org": "ByteDance", "country": "China", "date": "2025-08-20", "precision": "day", "category": "language", "open_weights": true, "params": "36B", "note": "Open 36B base and instruct models with 512K context", "source": "https://github.com/ByteDance-Seed/seed-oss" }, { "name": "Surya", "org": "IBM", "country": "USA", "date": "2025-08-20", "precision": "day", "category": "science", "open_weights": true, "params": "366M", "note": "IBM-NASA heliophysics foundation model forecasting solar flares", "source": "https://arxiv.org/abs/2508.14112" }, { "name": "Command A Reasoning", "org": "Cohere", "country": "Canada", "date": "2025-08-21", "precision": "day", "category": "reasoning", "open_weights": true, "params": "111B", "note": "Enterprise reasoning model with user-controlled token budget", "source": "https://cohere.com/blog/command-a-reasoning" }, { "name": "DeepSeek-V3.1", "org": "DeepSeek", "country": "China", "date": "2025-08-21", "precision": "day", "category": "language", "open_weights": true, "params": "671B (37B active)", "note": "Hybrid thinking/non-thinking model with stronger agent skills", "source": "https://api-docs.deepseek.com/news/news250821" }, { "name": "Grok 2.5", "org": "xAI", "country": "USA", "date": "2025-08-23", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "xAI released weights of its 2024 flagship model", "source": "https://techcrunch.com/2025/08/24/elon-musk-says-xai-has-open-sourced-grok-2-5/" }, { "name": "MiniCPM-V 4.5", "org": "ModelBest", "country": "China", "date": "2025-08-25", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Edge vision-language model claiming to beat GPT-4o-latest", "source": "https://github.com/OpenBMB/MiniCPM-V" }, { "name": "VibeVoice", "org": "Microsoft", "country": "USA", "date": "2025-08-25", "precision": "day", "category": "audio", "open_weights": true, "params": "1.5B", "note": "Long-form multi-speaker TTS generating up to 90 minutes", "source": "https://github.com/microsoft/VibeVoice" }, { "name": "Gemini 2.5 Flash Image (Nano Banana)", "org": "Google DeepMind", "country": "USA", "date": "2025-08-26", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Viral image-editing model with strong character consistency", "source": "https://developers.googleblog.com/en/introducing-gemini-2-5-flash-image/" }, { "name": "Hermes 4", "org": "Nous Research", "country": "USA", "date": "2025-08-26", "precision": "day", "category": "reasoning", "open_weights": true, "params": "405B", "note": "Hybrid reasoning open models in 14B, 70B and 405B sizes", "source": "https://huggingface.co/NousResearch/Hermes-4-405B" }, { "name": "InternVL3.5", "org": "Shanghai AI Lab", "country": "China", "date": "2025-08-26", "precision": "day", "category": "language", "open_weights": true, "params": "241B (28B active)", "note": "Open multimodal family with Cascade RL reasoning", "source": "https://github.com/OpenGVLab/InternVL" }, { "name": "OmniHuman-1.5", "org": "ByteDance", "country": "China", "date": "2025-08-26", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Avatar animation with cognitive simulation for expressive motion", "source": "https://arxiv.org/abs/2508.19209" }, { "name": "Wan2.2-S2V", "org": "Alibaba", "country": "China", "date": "2025-08-26", "precision": "day", "category": "video", "open_weights": true, "params": "14B", "note": "Audio-driven cinematic video generation model", "source": "https://github.com/Wan-Video/Wan2.2" }, { "name": "Command A Translate", "org": "Cohere", "country": "Canada", "date": "2025-08-28", "precision": "day", "category": "language", "open_weights": true, "params": "111B", "note": "Enterprise translation model for secure deployment", "source": "https://cohere.com/blog/command-a-translate" }, { "name": "gpt-realtime", "org": "OpenAI", "country": "USA", "date": "2025-08-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "OpenAI's most capable speech-to-speech model; Realtime API reached GA", "source": "https://developers.openai.com/api/docs/models/gpt-realtime" }, { "name": "Grok Code Fast 1", "org": "xAI", "country": "USA", "date": "2025-08-28", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Fast, economical reasoning model for agentic coding", "source": "https://x.ai/news/grok-code-fast-1" }, { "name": "HunyuanVideo-Foley", "org": "Tencent", "country": "China", "date": "2025-08-28", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open text-video-to-audio Foley sound generation", "source": "https://github.com/Tencent-Hunyuan/HunyuanVideo-Foley" }, { "name": "Kwai Keye-VL 1.5", "org": "Kuaishou", "country": "China", "date": "2025-08-28", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Slow-fast video encoding and long-CoT reasoning, 128K context", "source": "https://github.com/Kwai-Keye/Keye" }, { "name": "MAI-1-preview", "org": "Microsoft", "country": "USA", "date": "2025-08-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Microsoft AI's first in-house end-to-end trained foundation MoE", "source": "https://microsoft.ai/news/two-new-in-house-models/" }, { "name": "MAI-Voice-1", "org": "Microsoft", "country": "USA", "date": "2025-08-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Microsoft AI's first in-house expressive speech generation model", "source": "https://microsoft.ai/news/two-new-in-house-models/" }, { "name": "OLMoASR", "org": "Allen Institute for AI", "country": "USA", "date": "2025-08-28", "precision": "day", "category": "audio", "open_weights": true, "params": "1.5B", "note": "Fully open speech recognition models rivaling Whisper", "source": "https://allenai.org/blog/olmoasr" }, { "name": "PixVerse V5", "org": "PixVerse", "country": "China", "date": "2025-08-28", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Fifth-generation PixVerse video model", "source": "https://www.manilatimes.net/2025/08/28/tmt-newswire/pr-newswire/pixverse-launches-ai-video-model-v5-with-free-access-week/2175342" }, { "name": "YandexGPT 5.1 Pro", "org": "Yandex", "country": "Russia", "date": "2025-08-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Updated Yandex flagship LLM, released for testing", "source": "https://aistudio.yandex.ru/docs/en/ai-studio/release-notes/" }, { "name": "Step-Audio 2 mini", "org": "StepFun", "country": "China", "date": "2025-08-29", "precision": "day", "category": "audio", "open_weights": true, "params": "8B", "note": "Open end-to-end speech understanding and conversation model", "source": "https://github.com/stepfun-ai/Step-Audio2" }, { "name": "LongCat-Flash-Chat", "org": "Meituan", "country": "China", "date": "2025-08-30", "precision": "day", "category": "language", "open_weights": true, "params": "560B (~27B active)", "note": "Food-delivery giant's open MoE with dynamic zero-computation experts", "source": "https://huggingface.co/meituan-longcat/LongCat-Flash-Chat" }, { "name": "Holo1.5", "org": "H Company", "country": "France", "date": "2025-09-01", "precision": "month", "category": "agent", "open_weights": true, "params": "3B / 7B / 72B", "note": "Computer-use UI localization and screen VQA models", "source": "https://huggingface.co/Hcompany/Holo1.5-7B/commits/main" }, { "name": "Hunyuan-MT-7B", "org": "Tencent", "country": "China", "date": "2025-09-01", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Open translation model; first in 30 of 31 WMT25 categories", "source": "https://github.com/Tencent-Hunyuan/Hunyuan-MT" }, { "name": "Suno v5", "org": "Suno", "country": "USA", "date": "2025-09-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Suno's most advanced music model, launched alongside Suno Studio DAW", "source": "https://www.musicbusinessworldwide.com/suno-launches-its-own-daw-after-introducing-most-powerful-model-yet/" }, { "name": "VoxCPM", "org": "ModelBest", "country": "China", "date": "2025-09-01", "precision": "month", "category": "audio", "open_weights": true, "params": "0.5B", "note": "Tokenizer-free open TTS with context-aware voice cloning", "source": "https://github.com/OpenBMB/VoxCPM" }, { "name": "Apertus", "org": "Swiss AI Initiative", "country": "Switzerland", "date": "2025-09-02", "precision": "day", "category": "language", "open_weights": true, "params": "8B / 70B", "note": "Fully open Swiss LLM trained on 1,000+ languages", "source": "https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html" }, { "name": "HunyuanWorld-Voyager", "org": "Tencent", "country": "China", "date": "2025-09-02", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "RGB-D video diffusion for explorable 3D scenes from one image", "source": "https://github.com/Tencent-Hunyuan/HunyuanWorld-Voyager" }, { "name": "UI-TARS-2", "org": "ByteDance", "country": "China", "date": "2025-09-02", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "GUI agent trained with multi-turn RL for computer, game and tool use", "source": "https://arxiv.org/abs/2509.02544" }, { "name": "Chatterbox Multilingual", "org": "Resemble AI", "country": "USA", "date": "2025-09-04", "precision": "day", "category": "audio", "open_weights": true, "params": "500M", "note": "Open TTS with voice cloning in 23 languages", "source": "https://www.resemble.ai/introducing-chatterbox-multilingual-open-source-tts-for-23-languages/" }, { "name": "Deep Loop Shaping", "org": "Google DeepMind", "country": "USA", "date": "2025-09-04", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "RL controller cut LIGO mirror control noise 30-100x", "source": "https://deepmind.google/blog/using-ai-to-perceive-the-universe-in-greater-depth/" }, { "name": "EmbeddingGemma", "org": "Google DeepMind", "country": "USA", "date": "2025-09-04", "precision": "day", "category": "language", "open_weights": true, "params": "308M", "note": "Best-in-class small open embedding model for on-device use", "source": "https://developers.googleblog.com/en/introducing-embeddinggemma/" }, { "name": "Kimi-K2-Instruct-0905", "org": "Moonshot AI", "country": "China", "date": "2025-09-05", "precision": "day", "category": "language", "open_weights": true, "params": "1T (32B active)", "note": "Better agentic coding; context doubled to 256K", "source": "https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905" }, { "name": "MiniCPM4.1", "org": "ModelBest", "country": "China", "date": "2025-09-05", "precision": "day", "category": "reasoning", "open_weights": true, "params": "8B", "note": "Trainable sparse attention with hybrid reasoning for edge devices", "source": "https://github.com/OpenBMB/MiniCPM" }, { "name": "Qwen3-Max-Preview", "org": "Alibaba", "country": "China", "date": "2025-09-05", "precision": "day", "category": "language", "open_weights": false, "params": ">1T", "note": "First trillion-parameter Qwen, released as preview", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "HunyuanImage 2.1", "org": "Tencent", "country": "China", "date": "2025-09-08", "precision": "day", "category": "image", "open_weights": true, "params": "17B", "note": "Open 2K-resolution text-to-image model", "source": "https://github.com/Tencent-Hunyuan/HunyuanImage-2.1" }, { "name": "Qwen3-ASR-Flash", "org": "Alibaba", "country": "China", "date": "2025-09-08", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "LLM-based multilingual speech recognition with auto language detection", "source": "https://help.aliyun.com/zh/model-studio/newly-released-models" }, { "name": "ERNIE X1.1", "org": "Baidu", "country": "China", "date": "2025-09-09", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Upgraded reasoning model unveiled at WAVE SUMMIT 2025", "source": "https://www.prnewswire.com/news-releases/baidu-unveils-reasoning-model-ernie-x1-1-with-upgrades-in-key-capabilities-302551170.html" }, { "name": "ERNIE-4.5-21B-A3B-Thinking", "org": "Baidu", "country": "China", "date": "2025-09-09", "precision": "day", "category": "reasoning", "open_weights": true, "params": "21B (3B active)", "note": "Compact open MoE reasoning model under Apache 2.0", "source": "https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking" }, { "name": "K2 Think", "org": "MBZUAI", "country": "UAE", "date": "2025-09-09", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32B", "note": "Parameter-efficient open reasoning system from Abu Dhabi, with G42", "source": "https://arxiv.org/abs/2509.07604" }, { "name": "Seedream 4.0", "org": "ByteDance", "country": "China", "date": "2025-09-09", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Unified image generation and editing; rival to Nano Banana", "source": "https://seed.bytedance.com/blog/seedream-4-0-officially-released-beyond-drawing-into-imagination" }, { "name": "Stable Audio 2.5", "org": "Stability AI", "country": "UK", "date": "2025-09-10", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Enterprise audio generation model for sound production", "source": "https://stability.ai/news/stability-ai-introduces-stable-audio-25-the-first-audio-model-built-for-enterprise-sound-production-at-scale" }, { "name": "MiniMax Music 1.5", "org": "MiniMax", "country": "China", "date": "2025-09-11", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Four-minute full songs with improved vocals", "source": "https://www.minimax.io/news/minimax-music-15" }, { "name": "Qwen3-Next", "org": "Alibaba", "country": "China", "date": "2025-09-11", "precision": "day", "category": "language", "open_weights": true, "params": "80B (3B active)", "note": "Hybrid Gated DeltaNet attention with ultra-sparse MoE", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "MobileLLM-R1", "org": "Meta", "country": "USA", "date": "2025-09-12", "precision": "day", "category": "reasoning", "open_weights": true, "params": "140M-950M", "note": "Sub-billion-parameter reasoning models for edge devices", "source": "https://huggingface.co/facebook/MobileLLM-R1-950M" }, { "name": "VaultGemma", "org": "Google DeepMind", "country": "USA", "date": "2025-09-12", "precision": "day", "category": "language", "open_weights": true, "params": "1B", "note": "Largest open LLM trained from scratch with differential privacy", "source": "https://research.google/blog/vaultgemma-the-worlds-most-capable-differentially-private-llm/" }, { "name": "Fabric 1.0", "org": "Veed", "country": "UK", "date": "2025-09-15", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Talking-video model animating still images from audio", "source": "https://www.veed.io/ai/fabric-1-0" }, { "name": "GPT-5-Codex", "org": "OpenAI", "country": "USA", "date": "2025-09-15", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "GPT-5 variant optimized for agentic coding in Codex", "source": "https://openai.com/index/introducing-upgrades-to-codex/" }, { "name": "TimesFM 2.5", "org": "Google", "country": "USA", "date": "2025-09-15", "precision": "day", "category": "science", "open_weights": true, "params": "200M", "note": "Smaller time-series foundation model with 16K context", "source": "https://github.com/google-research/timesfm" }, { "name": "Hunyuan3D 3.0", "org": "Tencent", "country": "China", "date": "2025-09-16", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "3D generation model with tripled modeling precision", "source": "https://pandaily.com/tencent-unveils-hunyuan-3d-3-0-ai-model-tripling-modeling-accuracy-with-free-access" }, { "name": "Granite-Docling", "org": "IBM", "country": "USA", "date": "2025-09-17", "precision": "day", "category": "vision", "open_weights": true, "params": "258M", "note": "Tiny end-to-end document conversion vision-language model", "source": "https://www.ibm.com/new/announcements/granite-docling-end-to-end-document-conversion" }, { "name": "Ling-flash-2.0", "org": "Ant Group", "country": "China", "date": "2025-09-17", "precision": "day", "category": "language", "open_weights": true, "params": "100B (6.1B active)", "note": "Ling 2.0 MoE family flagship at launch", "source": "https://huggingface.co/inclusionAI/Ling-flash-2.0" }, { "name": "Magistral 1.2", "org": "Mistral AI", "country": "France", "date": "2025-09-17", "precision": "day", "category": "reasoning", "open_weights": true, "params": "24B", "note": "Magistral Small/Medium reasoning update adding vision input", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "Tongyi DeepResearch", "org": "Alibaba", "country": "China", "date": "2025-09-17", "precision": "day", "category": "agent", "open_weights": true, "params": "30B (3B active)", "note": "Open web deep-research agent model from Tongyi Lab", "source": "https://github.com/Alibaba-NLP/DeepResearch" }, { "name": "Moondream 3 Preview", "org": "Moondream", "country": "USA", "date": "2025-09-18", "precision": "day", "category": "vision", "open_weights": true, "params": "9B (2B active)", "note": "Small MoE vision model with grounded visual reasoning", "source": "https://moondream.ai/blog/moondream-3-preview" }, { "name": "Ray3", "org": "Luma AI", "country": "USA", "date": "2025-09-18", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Billed as first reasoning video model; native 16-bit HDR", "source": "https://lumalabs.ai/news/ray3" }, { "name": "RFdiffusion3", "org": "University of Washington", "country": "USA", "date": "2025-09-18", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "All-atom generative design of proteins with ligands and nucleic acids", "source": "https://www.biorxiv.org/content/10.1101/2025.09.18.676967v1" }, { "name": "Grok 4 Fast", "org": "xAI", "country": "USA", "date": "2025-09-19", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Cost-efficient unified reasoning model with 2M-token context", "source": "https://x.ai/news/grok-4-fast" }, { "name": "Manzano", "org": "Apple", "country": "USA", "date": "2025-09-19", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Unified image understanding and generation via hybrid tokenizer", "source": "https://arxiv.org/abs/2509.16197" }, { "name": "MiMo-Audio", "org": "Xiaomi", "country": "China", "date": "2025-09-19", "precision": "day", "category": "audio", "open_weights": true, "params": "7B", "note": "Speech LM trained on 100M+ hours showing few-shot audio abilities", "source": "https://www.marktechpost.com/2025/09/20/xiaomi-released-mimo-audio-a-7b-speech-language-model-trained-on-100m-hours-with-high-fidelity-discrete-tokens/" }, { "name": "MinerU2.5", "org": "Shanghai AI Lab", "country": "China", "date": "2025-09-19", "precision": "day", "category": "vision", "open_weights": true, "params": "1.2B", "note": "Decoupled document-parsing VLM from OpenDataLab beating far larger models", "source": "https://github.com/opendatalab/MinerU/releases/tag/mineru-2.5.0-released" }, { "name": "Wan2.2-Animate", "org": "Alibaba", "country": "China", "date": "2025-09-19", "precision": "day", "category": "video", "open_weights": true, "params": "14B", "note": "Unified character animation and replacement model", "source": "https://github.com/Wan-Video/Wan2.2" }, { "name": "DeepSeek-V3.1-Terminus", "org": "DeepSeek", "country": "China", "date": "2025-09-22", "precision": "day", "category": "language", "open_weights": true, "params": "671B (37B active)", "note": "V3.1 refresh fixing language mixing, stronger agents", "source": "https://api-docs.deepseek.com/news/news250922" }, { "name": "LongCat-Flash-Thinking", "org": "Meituan", "country": "China", "date": "2025-09-22", "precision": "day", "category": "reasoning", "open_weights": true, "params": "560B (~27B active)", "note": "Open reasoning model rivaling GPT-5 on some benchmarks", "source": "https://arxiv.org/abs/2509.18883" }, { "name": "Qwen-Image-Edit-2509", "org": "Alibaba", "country": "China", "date": "2025-09-22", "precision": "day", "category": "image", "open_weights": true, "params": "20B", "note": "Edit update adding multi-image editing and identity preservation", "source": "https://huggingface.co/Qwen/Qwen-Image-Edit-2509" }, { "name": "Qwen3-LiveTranslate-Flash", "org": "Alibaba", "country": "China", "date": "2025-09-22", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Real-time multilingual audio-video simultaneous interpretation model", "source": "https://help.aliyun.com/zh/model-studio/newly-released-models" }, { "name": "Qwen3-Omni", "org": "Alibaba", "country": "China", "date": "2025-09-22", "precision": "day", "category": "multimodal", "open_weights": true, "params": "30B (3B active)", "note": "Natively end-to-end omni model for text, image, audio and video", "source": "https://github.com/QwenLM/Qwen3-Omni" }, { "name": "Kling 2.5 Turbo", "org": "Kuaishou", "country": "China", "date": "2025-09-23", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Topped Artificial Analysis video arena at lower cost", "source": "https://blog.fal.ai/kling-2-5-turbo-pro-is-now-available-on-fal/" }, { "name": "LFM2-2.6B", "org": "Liquid AI", "country": "USA", "date": "2025-09-23", "precision": "day", "category": "language", "open_weights": true, "params": "2.6B", "note": "Largest dense LFM2 on-device model", "source": "https://www.liquid.ai/blog/introducing-lfm2-2-6b-redefining-efficiency-in-language-models" }, { "name": "Qwen3-VL", "org": "Alibaba", "country": "China", "date": "2025-09-23", "precision": "day", "category": "language", "open_weights": true, "params": "235B (22B active)", "note": "Flagship open vision-language model with visual agent abilities", "source": "https://github.com/QwenLM/Qwen3-VL" }, { "name": "Qwen3Guard", "org": "Alibaba", "country": "China", "date": "2025-09-23", "precision": "day", "category": "language", "open_weights": true, "params": "0.6B-8B", "note": "First Qwen safety guardrail model with streaming moderation", "source": "https://qwenlm.github.io/blog/" }, { "name": "SimpleFold", "org": "Apple", "country": "USA", "date": "2025-09-23", "precision": "day", "category": "science", "open_weights": true, "params": "3B", "note": "Protein folding with general-purpose transformers, no domain-specific modules", "source": "https://arxiv.org/abs/2509.18480" }, { "name": "Code World Model (CWM)", "org": "Meta", "country": "USA", "date": "2025-09-24", "precision": "day", "category": "code", "open_weights": true, "params": "32B", "note": "Open research LLM for code generation trained on execution traces", "source": "https://huggingface.co/facebook/cwm" }, { "name": "Qwen3-Max", "org": "Alibaba", "country": "China", "date": "2025-09-24", "precision": "day", "category": "language", "open_weights": false, "params": ">1T", "note": "Official trillion-parameter flagship launched at Apsara Conference", "source": "https://en.wikipedia.org/wiki/Qwen" }, { "name": "Wan2.5-Preview", "org": "Alibaba", "country": "China", "date": "2025-09-24", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Adds synchronized audio to 10-second generated videos", "source": "https://blog.fal.ai/wan-2-5-preview-is-now-available-on-fal/" }, { "name": "Gemini 2.5 Flash Preview 09-2025", "org": "Google DeepMind", "country": "USA", "date": "2025-09-25", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Improved Flash and Flash-Lite previews with better agentic tool use", "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/" }, { "name": "Gemini Robotics 1.5", "org": "Google DeepMind", "country": "USA", "date": "2025-09-25", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Vision-language-action model that thinks before acting", "source": "https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/" }, { "name": "Gemini Robotics-ER 1.5", "org": "Google DeepMind", "country": "USA", "date": "2025-09-25", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Embodied reasoning model for robot planning, available via API", "source": "https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/" }, { "name": "ShinkaEvolve", "org": "Sakana AI", "country": "Japan", "date": "2025-09-25", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Open-source, sample-efficient LLM-driven program evolution framework", "source": "https://sakana.ai/blog/" }, { "name": "Vidu Q2", "org": "ShengShu", "country": "China", "date": "2025-09-25", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Image-to-video model focused on AI acting and facial performance", "source": "https://www.geekpark.net/news/354451" }, { "name": "HunyuanImage 3.0", "org": "Tencent", "country": "China", "date": "2025-09-28", "precision": "day", "category": "image", "open_weights": true, "params": "80B (13B active)", "note": "Largest open-source image generation MoE model at release", "source": "https://github.com/Tencent-Hunyuan/HunyuanImage-3.0" }, { "name": "Claude Sonnet 4.5", "org": "Anthropic", "country": "USA", "date": "2025-09-29", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Anthropic's claimed best coding model and strongest for building agents", "source": "https://www.anthropic.com/news/claude-sonnet-4-5" }, { "name": "Cosmos Predict 2.5", "org": "NVIDIA", "country": "USA", "date": "2025-09-29", "precision": "day", "category": "video", "open_weights": true, "params": "2B / 14B", "note": "Unified text/image/video-to-world model, released with Transfer 2.5", "source": "https://nvidianews.nvidia.com/news/nvidia-accelerates-robotics-research-and-development-with-new-open-models-and-simulation-libraries" }, { "name": "DeepSeek-V3.2-Exp", "org": "DeepSeek", "country": "China", "date": "2025-09-29", "precision": "day", "category": "language", "open_weights": true, "params": "671B (37B active)", "note": "Debuted DeepSeek Sparse Attention; API prices cut by over 50%", "source": "https://api-docs.deepseek.com/news/news250929" }, { "name": "Isaac GR00T N1.6", "org": "NVIDIA", "country": "USA", "date": "2025-09-29", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Open humanoid VLA model using Cosmos Reason for reasoning", "source": "https://nvidianews.nvidia.com/news/nvidia-accelerates-robotics-research-and-development-with-new-open-models-and-simulation-libraries" }, { "name": "Kandinsky 5.0 Video Lite", "org": "Sber", "country": "Russia", "date": "2025-09-29", "precision": "day", "category": "video", "open_weights": true, "params": "2B", "note": "Open lightweight text-to-video model, up to 10-second clips", "source": "https://github.com/kandinskylab/kandinsky-5" }, { "name": "RoboBrain-X0", "org": "BAAI", "country": "China", "date": "2025-09-29", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Open cross-embodiment VLA with zero-shot generalization across robots", "source": "https://github.com/FlagOpen/RoboBrain-X0" }, { "name": "GLM-4.6", "org": "Zhipu AI", "country": "China", "date": "2025-09-30", "precision": "day", "category": "language", "open_weights": true, "params": "355B (32B active)", "note": "Coding-focused GLM upgrade with 200K context", "source": "https://z.ai/blog/glm-4.6" }, { "name": "Sora 2", "org": "OpenAI", "country": "USA", "date": "2025-09-30", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Video-with-audio model launched with TikTok-style Sora app", "source": "https://openai.com/index/sora-2/" }, { "name": "Sora 2 Pro", "org": "OpenAI", "country": "USA", "date": "2025-09-30", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Higher-quality Sora 2 tier for ChatGPT Pro users, API from DevDay", "source": "https://openai.com/index/sora-2/" }, { "name": "Apriel-1.5-15B-Thinker", "org": "ServiceNow", "country": "USA", "date": "2025-10-01", "precision": "day", "category": "reasoning", "open_weights": true, "params": "15B", "note": "15B open multimodal reasoner scoring near much larger frontier models", "source": "https://huggingface.co/ServiceNow-AI/Apriel-1.5-15b-Thinker" }, { "name": "FIBO", "org": "Bria AI", "country": "Israel", "date": "2025-10-01", "precision": "month", "category": "image", "open_weights": true, "params": "8B", "note": "First open JSON-native text-to-image model trained on licensed data", "source": "https://huggingface.co/briaai/FIBO" }, { "name": "Lapa LLM", "org": "Lapa LLM (Ukrainian universities consortium)", "country": "Ukraine", "date": "2025-10-01", "precision": "month", "category": "language", "open_weights": true, "params": "12B", "note": "Ukraine's open Gemma-3-based LLM with Ukrainian-optimized tokenizer", "source": "https://huggingface.co/lapa-llm/lapa-v0.1.2-instruct" }, { "name": "LFM2-Audio", "org": "Liquid AI", "country": "USA", "date": "2025-10-01", "precision": "day", "category": "audio", "open_weights": true, "params": "1.5B", "note": "End-to-end on-device audio foundation model", "source": "https://www.liquid.ai/blog/lfm2-audio-an-end-to-end-audio-foundation-model" }, { "name": "Octave 2", "org": "Hume AI", "country": "USA", "date": "2025-10-01", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Next-generation multilingual expressive voice AI model", "source": "https://www.hume.ai/blog/octave-2-launch" }, { "name": "PLaMo 3 NICT", "org": "Preferred Networks", "country": "Japan", "date": "2025-10-01", "precision": "month", "category": "language", "open_weights": true, "params": "2B / 8B / 31B", "note": "Japanese-English hybrid sliding-window base models co-developed with NICT", "source": "https://huggingface.co/pfnet/plamo-3-nict-31b-base" }, { "name": "Suno v4.5-all", "org": "Suno", "country": "USA", "date": "2025-10-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "New default free-tier model for all Suno users", "source": "https://en.wikipedia.org/wiki/Suno_AI" }, { "name": "Granite 4.0", "org": "IBM", "country": "USA", "date": "2025-10-02", "precision": "day", "category": "language", "open_weights": true, "params": "32B (9B active)", "note": "Hyper-efficient hybrid Mamba/Transformer enterprise models", "source": "https://www.ibm.com/new/announcements/ibm-granite-4-0-hyper-efficient-high-performance-hybrid-models" }, { "name": "CodeMender", "org": "Google DeepMind", "country": "USA", "date": "2025-10-06", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "AI agent that finds and patches code security vulnerabilities", "source": "https://deepmind.google/blog/introducing-codemender-an-ai-agent-for-code-security/" }, { "name": "gpt-image-1-mini", "org": "OpenAI", "country": "USA", "date": "2025-10-06", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Lower-cost image generation and editing model from DevDay", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "gpt-realtime-mini", "org": "OpenAI", "country": "USA", "date": "2025-10-06", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Cheaper speech-to-speech model from DevDay, with gpt-audio-mini", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "Tiny Recursive Model (TRM)", "org": "Samsung", "country": "South Korea", "date": "2025-10-06", "precision": "day", "category": "reasoning", "open_weights": true, "params": "7M", "note": "7M-parameter recursive network scoring 45% on ARC-AGI-1", "source": "https://arxiv.org/abs/2510.04871" }, { "name": "Gemini 2.5 Computer Use", "org": "Google DeepMind", "country": "USA", "date": "2025-10-07", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Model that operates browser and mobile UIs by clicking and typing", "source": "https://blog.google/technology/google-deepmind/gemini-computer-use-model/" }, { "name": "LFM2-8B-A1B", "org": "Liquid AI", "country": "USA", "date": "2025-10-07", "precision": "day", "category": "language", "open_weights": true, "params": "8.3B (1.5B active)", "note": "Liquid AI's first on-device mixture-of-experts model", "source": "https://www.liquid.ai/blog/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts" }, { "name": "Jamba Reasoning 3B", "org": "AI21 Labs", "country": "Israel", "date": "2025-10-08", "precision": "day", "category": "reasoning", "open_weights": true, "params": "3B", "note": "Tiny SSM-Transformer reasoner with 256K context window", "source": "https://www.ai21.com/blog/introducing-jamba-reasoning-3b/" }, { "name": "Ling-1T", "org": "Ant Group", "country": "China", "date": "2025-10-09", "precision": "day", "category": "language", "open_weights": true, "params": "1T (~50B active)", "note": "Open trillion-parameter non-thinking flagship", "source": "https://www.businesswire.com/news/home/20251009240721/en/Ant-Group-Unveils-Ling-AI-Model-Family-and-Launches-Trillion-Parameter-Language-Model-Ling-1T" }, { "name": "MAI-Image-1", "org": "Microsoft", "country": "USA", "date": "2025-10-13", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Microsoft AI's first in-house image generator, debuted top-10 on LMArena", "source": "https://microsoft.ai/news/introducing-mai-image-1-debuting-in-the-top-10-on-lmarena/" }, { "name": "Ring-1T", "org": "Ant Group", "country": "China", "date": "2025-10-14", "precision": "day", "category": "reasoning", "open_weights": true, "params": "1T (50B active)", "note": "First open trillion-parameter thinking model; preview came Sep 30", "source": "https://huggingface.co/inclusionAI/Ring-1T" }, { "name": "C2S-Scale 27B", "org": "Google DeepMind", "country": "USA", "date": "2025-10-15", "precision": "day", "category": "science", "open_weights": true, "params": "27B", "note": "Gemma-based single-cell model surfaced a validated cancer therapy hypothesis", "source": "https://blog.google/technology/ai/google-gemma-ai-cancer-therapy-discovery/" }, { "name": "Claude Haiku 4.5", "org": "Anthropic", "country": "USA", "date": "2025-10-15", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Sonnet 4-level coding at one-third the cost, twice the speed", "source": "https://www.anthropic.com/news/claude-haiku-4-5" }, { "name": "Odyssey", "org": "Anthrogen", "country": "USA", "date": "2025-10-15", "precision": "day", "category": "science", "open_weights": false, "params": "102B", "note": "Protein language models up to 102B parameters", "source": "https://www.biorxiv.org/content/10.1101/2025.10.15.682677v1" }, { "name": "Qwen3-VL-8B", "org": "Alibaba", "country": "China", "date": "2025-10-15", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Compact dense Qwen3-VL models (4B/8B) for edge multimodal use", "source": "https://github.com/QwenLM/Qwen3-VL" }, { "name": "Veo 3.1", "org": "Google DeepMind", "country": "USA", "date": "2025-10-15", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Veo update with richer audio, reference images and scene extension", "source": "https://blog.google/technology/ai/veo-updates-flow/" }, { "name": "DeepSomatic", "org": "Google", "country": "USA", "date": "2025-10-16", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Identifies cancer-related genetic variants in tumor sequencing", "source": "https://research.google/blog/using-ai-to-identify-genetic-variants-in-tumors-with-deepsomatic/" }, { "name": "PaddleOCR-VL", "org": "Baidu", "country": "China", "date": "2025-10-16", "precision": "day", "category": "vision", "open_weights": true, "params": "0.9B", "note": "Ultra-compact multilingual document parsing vision-language model", "source": "https://arxiv.org/abs/2510.14528" }, { "name": "RTFM", "org": "World Labs", "country": "USA", "date": "2025-10-16", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Real-time frame world model rendering 3D worlds on a single H100", "source": "https://www.worldlabs.ai/blog/rtfm" }, { "name": "SWE-grep", "org": "Cognition", "country": "USA", "date": "2025-10-16", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "RL-trained fast multi-turn context retrieval for coding agents", "source": "https://cognition.ai/blog/swe-grep" }, { "name": "Chronos-2", "org": "Amazon", "country": "USA", "date": "2025-10-20", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Universal zero-shot time-series forecasting with covariate support", "source": "https://www.amazon.science/blog/introducing-chronos-2-from-univariate-to-universal-forecasting" }, { "name": "DeepSeek-OCR", "org": "DeepSeek", "country": "China", "date": "2025-10-20", "precision": "day", "category": "vision", "open_weights": true, "params": "3B", "note": "Compresses long text into vision tokens via optical compression", "source": "https://arxiv.org/abs/2510.18234" }, { "name": "Krea Realtime 14B", "org": "Krea", "country": "USA", "date": "2025-10-20", "precision": "day", "category": "video", "open_weights": true, "params": "14B", "note": "Open real-time long-form autoregressive video generation model", "source": "https://www.krea.ai/blog/krea-realtime-14b" }, { "name": "tsuzumi 2", "org": "NTT", "country": "Japan", "date": "2025-10-20", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Japanese enterprise LLM that runs on a single GPU", "source": "https://group.ntt/en/newsrelease/2025/10/20/251020a.html" }, { "name": "Baichuan-M2 Plus", "org": "Baichuan", "country": "China", "date": "2025-10-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Evidence-augmented medical model billed as ChatGPT for doctors", "source": "https://pandaily.com/baichuan-releases-m2-plus-an-evidence-augmented-medical-model-billed-as-a-chat-gpt-for-doctors" }, { "name": "LFM2-VL-3B", "org": "Liquid AI", "country": "USA", "date": "2025-10-22", "precision": "day", "category": "language", "open_weights": true, "params": "3B", "note": "Larger efficient edge vision-language model", "source": "https://www.liquid.ai/blog/lfm2-vl-3b-a-new-efficient-vision-language-for-the-edge" }, { "name": "olmOCR 2", "org": "Allen Institute for AI", "country": "USA", "date": "2025-10-22", "precision": "day", "category": "vision", "open_weights": true, "params": "7B", "note": "Document OCR model trained with unit-test rewards", "source": "https://allenai.org/blog/olmocr-2" }, { "name": "LTX-2", "org": "Lightricks", "country": "Israel", "date": "2025-10-23", "precision": "day", "category": "video", "open_weights": true, "params": "19B", "note": "Synchronized 4K audio-video generation; weights opened January 2026", "source": "https://www.ynetnews.com/tech-and-digital/article/hklbzavrgx" }, { "name": "Seed3D 1.0", "org": "ByteDance", "country": "China", "date": "2025-10-23", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Single image to simulation-ready 3D assets", "source": "https://seed.bytedance.com/en/seed3d" }, { "name": "LongCat-Video", "org": "Meituan", "country": "China", "date": "2025-10-25", "precision": "day", "category": "video", "open_weights": true, "params": "13.6B", "note": "Open foundational video generation model for long videos", "source": "https://github.com/meituan-longcat/LongCat-Video" }, { "name": "BoltzGen", "org": "Boltz", "country": "USA", "date": "2025-10-26", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open generative model for universal protein binder design", "source": "https://boltz.com/news" }, { "name": "MiniMax-M2", "org": "MiniMax", "country": "China", "date": "2025-10-27", "precision": "day", "category": "language", "open_weights": true, "params": "230B (10B active)", "note": "Open agentic coding MoE priced at 8% of Claude Sonnet", "source": "https://www.minimax.io/news/minimax-m2" }, { "name": "Alice AI LLM", "org": "Yandex", "country": "Russia", "date": "2025-10-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "YandexGPT line relaunched as Alice AI model family", "source": "https://en.wikipedia.org/wiki/Alice_AI_(AI_model_family)" }, { "name": "Amazon Nova Multimodal Embeddings", "org": "Amazon", "country": "USA", "date": "2025-10-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Single embedding model for text, documents, images, video and audio", "source": "https://aws.amazon.com/blogs/aws/amazon-nova-multimodal-embeddings-now-available-in-amazon-bedrock/" }, { "name": "Firefly Image Model 5", "org": "Adobe", "country": "USA", "date": "2025-10-28", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Native 4MP photorealistic image model unveiled at Adobe MAX", "source": "https://news.adobe.com/news/2025/10/adobe-max-2025-firefly" }, { "name": "Granite 4.0 Nano", "org": "IBM", "country": "USA", "date": "2025-10-28", "precision": "day", "category": "language", "open_weights": true, "params": "350M-1.5B", "note": "Tiny hybrid-SSM models for edge and on-device use", "source": "https://huggingface.co/blog/ibm-granite/granite-4-nano" }, { "name": "Hailuo 2.3", "org": "MiniMax", "country": "China", "date": "2025-10-28", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Better body motion and micro-expressions, plus a Fast variant", "source": "https://www.minimax.io/news/minimax-hailuo-23" }, { "name": "LFM2-ColBERT-350M", "org": "Liquid AI", "country": "USA", "date": "2025-10-28", "precision": "day", "category": "language", "open_weights": true, "params": "350M", "note": "Multilingual late-interaction retrieval embedding model", "source": "https://www.liquid.ai/blog/lfm2-colbert-350m-one-model-to-embed-them-all" }, { "name": "Nemotron Nano 2 VL", "org": "NVIDIA", "country": "USA", "date": "2025-10-28", "precision": "day", "category": "language", "open_weights": true, "params": "12B", "note": "Open document and video intelligence vision-language model", "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16" }, { "name": "Composer", "org": "Cursor", "country": "USA", "date": "2025-10-29", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Cursor's first in-house fast frontier coding model trained with RL", "source": "https://cursor.com/blog/composer" }, { "name": "gpt-oss-safeguard", "org": "OpenAI", "country": "USA", "date": "2025-10-29", "precision": "day", "category": "reasoning", "open_weights": true, "params": "117B (5.1B active) / 21B (3.6B active)", "note": "Open safety-reasoning models (120b, 20b) following developer-written policies", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "MiniMax Music 2.0", "org": "MiniMax", "country": "China", "date": "2025-10-29", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music model with enhanced emotional expression", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "MiniMax Speech 2.6", "org": "MiniMax", "country": "China", "date": "2025-10-29", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Low-latency speech model built for real-time voice agents", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "SWE-1.5", "org": "Cognition", "country": "USA", "date": "2025-10-29", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Fast agentic coding model served at up to 950 tokens/s", "source": "https://cognition.ai/blog/swe-1-5" }, { "name": "Alpamayo-R1", "org": "NVIDIA", "country": "USA", "date": "2025-10-30", "precision": "day", "category": "robotics", "open_weights": true, "params": "10B", "note": "Reasoning VLA for driving; renamed Alpamayo 1 at CES 2026", "source": "https://arxiv.org/abs/2511.00088" }, { "name": "Emu3.5", "org": "BAAI", "country": "China", "date": "2025-10-30", "precision": "day", "category": "multimodal", "open_weights": true, "params": "34B", "note": "Native multimodal next-token model framed as a world learner", "source": "https://arxiv.org/abs/2510.26583" }, { "name": "Kimi Linear", "org": "Moonshot AI", "country": "China", "date": "2025-10-30", "precision": "day", "category": "language", "open_weights": true, "params": "48B (3B active)", "note": "Hybrid linear attention that beats full attention in fair tests", "source": "https://arxiv.org/abs/2510.26692" }, { "name": "LongCat-Flash-Omni", "org": "Meituan", "country": "China", "date": "2025-10-31", "precision": "day", "category": "multimodal", "open_weights": true, "params": "560B (27B active)", "note": "Open omni-modal model with real-time audio-visual interaction", "source": "https://arxiv.org/abs/2511.00279" }, { "name": "DeepL Agent", "org": "DeepL", "country": "Germany", "date": "2025-11-01", "precision": "month", "category": "agent", "open_weights": false, "params": "", "note": "AI agent that operates business applications", "source": "https://en.wikipedia.org/wiki/DeepL_Translator" }, { "name": "GigaChat 3 Ultra Preview", "org": "Sber", "country": "Russia", "date": "2025-11-01", "precision": "month", "category": "language", "open_weights": true, "params": "~700B", "note": "Sber's MIT-licensed open flagship MoE, plus 10B Lightning", "source": "https://habr.com/ru/companies/sberdevices/articles/" }, { "name": "GEN-0", "org": "Generalist AI", "country": "USA", "date": "2025-11-04", "precision": "day", "category": "robotics", "open_weights": false, "params": "10B", "note": "Embodied foundation model trained on 270,000+ hours of real manipulation", "source": "https://generalistai.com/blog/nov-04-2025-GEN-0" }, { "name": "OlmoEarth", "org": "Allen Institute for AI", "country": "USA", "date": "2025-11-04", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open Earth-observation foundation models and platform", "source": "https://allenai.org/blog/olmoearth" }, { "name": "Kimi K2 Thinking", "org": "Moonshot AI", "country": "China", "date": "2025-11-06", "precision": "day", "category": "reasoning", "open_weights": true, "params": "1T (32B active)", "note": "Open trillion-parameter thinking agent built for long-horizon tool use", "source": "https://moonshotai.github.io/Kimi-K2/thinking.html" }, { "name": "Spark X1.5", "org": "iFlytek", "country": "China", "date": "2025-11-06", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Deep reasoning model on domestic compute with voice cloning", "source": "https://pandaily.com/i-flytek-unveils-spark-x1-5-deep-reasoning-model-ai-that-truly-understands-you" }, { "name": "Step-Audio-EditX", "org": "StepFun", "country": "China", "date": "2025-11-06", "precision": "day", "category": "audio", "open_weights": true, "params": "3B", "note": "Open LLM-grade expressive audio editing model", "source": "https://github.com/stepfun-ai/Step-Audio-EditX" }, { "name": "Motif 2 12.7B", "org": "Motif Technologies", "country": "South Korea", "date": "2025-11-07", "precision": "day", "category": "language", "open_weights": true, "params": "12.7B", "note": "Korean startup Motif's open 12.7B language model", "source": "https://huggingface.co/Motif-Technologies/Motif-2-12.7B-Instruct" }, { "name": "GPT-5-Codex-Mini", "org": "OpenAI", "country": "USA", "date": "2025-11-08", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Smaller, cheaper GPT-5-Codex for Codex CLI usage", "source": "https://simonwillison.net/2025/Nov/9/gpt-5-codex-mini/" }, { "name": "Omnilingual ASR", "org": "Meta", "country": "USA", "date": "2025-11-10", "precision": "day", "category": "audio", "open_weights": true, "params": "7B", "note": "Speech recognition for 1,600+ languages, 500 never served before", "source": "https://ai.meta.com/blog/omnilingual-asr-advancing-automatic-speech-recognition/" }, { "name": "Doubao-Seed-Code", "org": "ByteDance", "country": "China", "date": "2025-11-11", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Agentic coding model priced for the Claude-cutoff market", "source": "https://eu.36kr.com/en/p/3548431703519367" }, { "name": "ERNIE-4.5-VL-28B-A3B-Thinking", "org": "Baidu", "country": "China", "date": "2025-11-11", "precision": "day", "category": "reasoning", "open_weights": true, "params": "28B (3B active)", "note": "Open vision-language reasoning model that 'thinks with images'", "source": "https://www.reddit.com/r/LocalLLaMA/comments/1ou14ry/baiduernie45vl28ba3bthinking_released_curious_case/" }, { "name": "Scribe v2 Realtime", "org": "ElevenLabs", "country": "USA", "date": "2025-11-11", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Low-latency real-time speech-to-text for voice agents", "source": "https://elevenlabs.io/blog/introducing-scribe-v2-realtime" }, { "name": "SenseNova-SI", "org": "SenseTime", "country": "China", "date": "2025-11-11", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Open spatial-intelligence model claimed to beat GPT-5 on spatial tasks", "source": "https://www.gadgets360.com/ai/news/sensetime-sensenova-si-open-source-ai-model-outperforms-chatgpt-gpt-5-gemini-2-5-pro-spatial-intelligence-9614833" }, { "name": "GPT-5.1", "org": "OpenAI", "country": "USA", "date": "2025-11-12", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Warmer, more steerable GPT-5 with Instant and Thinking modes", "source": "https://en.wikipedia.org/wiki/GPT-5.1" }, { "name": "Marble", "org": "World Labs", "country": "USA", "date": "2025-11-12", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Commercial multimodal world model generating editable 3D worlds", "source": "https://www.worldlabs.ai/blog" }, { "name": "PAN", "org": "MBZUAI", "country": "UAE", "date": "2025-11-12", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "General, actionable long-horizon world simulation model", "source": "https://arxiv.org/abs/2511.09057" }, { "name": "Depth Anything 3", "org": "ByteDance", "country": "China", "date": "2025-11-13", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Recovers 3D visual space from any number of views", "source": "https://www.alphaxiv.org/abs/2511.10647" }, { "name": "ERNIE 5.0", "org": "Baidu", "country": "China", "date": "2025-11-13", "precision": "day", "category": "multimodal", "open_weights": false, "params": "2.4T (MoE)", "note": "Natively omni-modal 2.4T MoE; preview Nov 2025, GA Jan 22, 2026", "source": "https://www.prnewswire.com/news-releases/baidu-unveils-ernie-5-0-and-a-series-of-ai-applications-at-baidu-world-2025--ramps-up-global-push-302614531.html" }, { "name": "GPT-5.1-Codex", "org": "OpenAI", "country": "USA", "date": "2025-11-13", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "GPT-5.1 variant optimized for agentic coding", "source": "https://platform.openai.com/docs/changelog" }, { "name": "GPT-5.1-Codex-Mini", "org": "OpenAI", "country": "USA", "date": "2025-11-13", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Smaller, cheaper GPT-5.1 coding model", "source": "https://platform.openai.com/docs/changelog" }, { "name": "Holo2", "org": "H Company", "country": "France", "date": "2025-11-13", "precision": "day", "category": "agent", "open_weights": true, "params": "4B / 8B / 30B", "note": "Computer-use agent models fine-tuned from Qwen3-VL", "source": "https://www.hcompany.ai/blog/holo2" }, { "name": "SIMA 2", "org": "Google DeepMind", "country": "USA", "date": "2025-11-13", "precision": "day", "category": "games", "open_weights": false, "params": "", "note": "Gemini-powered agent that plays, reasons and learns in 3D games", "source": "https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/" }, { "name": "Grok 4.1", "org": "xAI", "country": "USA", "date": "2025-11-17", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Fewer hallucinations, more emotional intelligence; #1 on LMArena text", "source": "https://x.ai/news/grok-4-1" }, { "name": "P1", "org": "Shanghai AI Lab", "country": "China", "date": "2025-11-17", "precision": "day", "category": "reasoning", "open_weights": true, "params": "235B (22B active)", "note": "First open model with gold-medal performance at IPhO 2025", "source": "https://arxiv.org/abs/2511.13612" }, { "name": "WeatherNext 2", "org": "Google DeepMind", "country": "USA", "date": "2025-11-17", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "More efficient, accurate, higher-resolution global weather forecasting model", "source": "https://blog.google/technology/google-deepmind/weathernext-2/" }, { "name": "π*0.6", "org": "Physical Intelligence", "country": "USA", "date": "2025-11-17", "precision": "day", "category": "robotics", "open_weights": false, "params": "5.3B", "note": "Physical Intelligence VLA model trained to learn from experience", "source": "https://www.pi.website/download/pistar06.pdf" }, { "name": "DR Tulu", "org": "Allen Institute for AI", "country": "USA", "date": "2025-11-18", "precision": "day", "category": "agent", "open_weights": true, "params": "8B", "note": "Open end-to-end training recipe for long-form deep research agents", "source": "https://allenai.org/blog/dr-tulu" }, { "name": "Gemini 3 Deep Think", "org": "Google DeepMind", "country": "USA", "date": "2025-11-18", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Enhanced reasoning mode of Gemini 3 for Ultra subscribers", "source": "https://blog.google/products/gemini/gemini-3/" }, { "name": "Gemini 3 Pro", "org": "Google DeepMind", "country": "USA", "date": "2025-11-18", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Google's most intelligent model; launched with Antigravity agentic IDE", "source": "https://blog.google/products/gemini/gemini-3/" }, { "name": "Cogito v2.1", "org": "Deep Cogito", "country": "USA", "date": "2025-11-19", "precision": "day", "category": "language", "open_weights": true, "params": "671B (37B active)", "note": "Open 671B hybrid reasoning model trained via iterated distillation", "source": "https://huggingface.co/deepcogito/cogito-671b-v2.1" }, { "name": "GPT-5.1 Pro", "org": "OpenAI", "country": "USA", "date": "2025-11-19", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Replaced GPT-5 Pro as the highest-compute ChatGPT model", "source": "https://en.wikipedia.org/wiki/GPT-5.1" }, { "name": "GPT-5.1-Codex-Max", "org": "OpenAI", "country": "USA", "date": "2025-11-19", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Long-horizon coding model using compaction across context windows", "source": "https://simonwillison.net/2025/Nov/19/gpt-51-codex-max/" }, { "name": "Grok 4.1 Fast", "org": "xAI", "country": "USA", "date": "2025-11-19", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Agentic tool-calling model launched with Agent Tools API", "source": "https://x.ai/news/grok-4-1-fast" }, { "name": "SAM 3", "org": "Meta", "country": "USA", "date": "2025-11-19", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Text-promptable concept segmentation and tracking", "source": "https://github.com/facebookresearch/sam3" }, { "name": "SAM 3D", "org": "Meta", "country": "USA", "date": "2025-11-19", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Reconstructs 3D objects and human bodies from a single image", "source": "https://github.com/facebookresearch/sam-3d-objects" }, { "name": "Step-Audio-R1", "org": "StepFun", "country": "China", "date": "2025-11-19", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Audio LLM that benefits from test-time reasoning", "source": "https://github.com/stepfun-ai/Step-Audio-R1" }, { "name": "HunyuanVideo 1.5", "org": "Tencent", "country": "China", "date": "2025-11-20", "precision": "day", "category": "video", "open_weights": true, "params": "8.3B", "note": "Lightweight open video model running on consumer GPUs", "source": "https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5" }, { "name": "Kandinsky 5.0 Video Pro", "org": "Sber", "country": "Russia", "date": "2025-11-20", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Ranked top open-source text-to-video model on LMArena", "source": "https://github.com/kandinskylab/kandinsky-5" }, { "name": "Keye-VL-671B-A37B", "org": "Kuaishou", "country": "China", "date": "2025-11-20", "precision": "day", "category": "language", "open_weights": true, "params": "671B (37B active)", "note": "Largest Keye multimodal model", "source": "https://github.com/Kwai-Keye/Keye" }, { "name": "MiMo-Embodied", "org": "Xiaomi", "country": "China", "date": "2025-11-20", "precision": "day", "category": "robotics", "open_weights": true, "params": "7B", "note": "Cross-embodied foundation model for embodied AI tasks", "source": "https://arxiv.org/abs/2511.16518" }, { "name": "Nano Banana Pro (Gemini 3 Pro Image)", "org": "Google DeepMind", "country": "USA", "date": "2025-11-20", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Gemini 3 Pro-based image model with legible multilingual text rendering", "source": "https://blog.google/technology/ai/nano-banana-pro/" }, { "name": "Olmo 3", "org": "Allen Institute for AI", "country": "USA", "date": "2025-11-20", "precision": "day", "category": "language", "open_weights": true, "params": "32B", "note": "Fully open models releasing the complete model flow, incl. Think variant", "source": "https://allenai.org/blog/olmo3" }, { "name": "ChatGPT shopping research model", "org": "OpenAI", "country": "USA", "date": "2025-11-24", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Mini model post-trained from GPT-5 thinking mini for shopping", "source": "https://en.wikipedia.org/wiki/GPT-5" }, { "name": "Claude Opus 4.5", "org": "Anthropic", "country": "USA", "date": "2025-11-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Frontier coding and agent model at one-third prior Opus price", "source": "https://www.anthropic.com/news/claude-opus-4-5" }, { "name": "Fara-7B", "org": "Microsoft", "country": "USA", "date": "2025-11-24", "precision": "day", "category": "agent", "open_weights": true, "params": "7B", "note": "Efficient open agentic model for computer use", "source": "https://www.microsoft.com/en-us/research/blog/fara-7b-an-efficient-agentic-model-for-computer-use/" }, { "name": "ZAYA1", "org": "Zyphra", "country": "USA", "date": "2025-11-24", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "MoE model pretrained on an integrated AMD GPU platform", "source": "https://www.zyphra.com/post/zaya1" }, { "name": "FLUX.2", "org": "Black Forest Labs", "country": "Germany", "date": "2025-11-25", "precision": "day", "category": "image", "open_weights": true, "params": "32B", "note": "Multi-reference image generation and editing; open 32B dev model", "source": "https://bfl.ai/blog/flux-2" }, { "name": "HunyuanOCR", "org": "Tencent", "country": "China", "date": "2025-11-25", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Lightweight end-to-end OCR vision-language model", "source": "https://github.com/Tencent-Hunyuan/HunyuanOCR" }, { "name": "INTELLECT-3", "org": "Prime Intellect", "country": "USA", "date": "2025-11-26", "precision": "day", "category": "reasoning", "open_weights": true, "params": "106B (12B active)", "note": "RL-trained on GLM-4.5-Air base with Prime Intellect's async PRIME-RL stack", "source": "https://www.primeintellect.ai/blog/intellect-3" }, { "name": "Z-Image-Turbo", "org": "Alibaba", "country": "China", "date": "2025-11-26", "precision": "day", "category": "image", "open_weights": true, "params": "6B", "note": "Efficient 6B open image model generating in 8 steps", "source": "https://github.com/Tongyi-MAI/Z-Image" }, { "name": "DeepSeekMath-V2", "org": "DeepSeek", "country": "China", "date": "2025-11-27", "precision": "day", "category": "reasoning", "open_weights": true, "params": "685B", "note": "Self-verifying math prover reaching IMO 2025 gold level", "source": "https://arxiv.org/abs/2511.22570" }, { "name": "DeepSeek-V3.2", "org": "DeepSeek", "country": "China", "date": "2025-12-01", "precision": "day", "category": "language", "open_weights": true, "params": "685B (37B active)", "note": "Official V3.2 with sparse attention and thinking integrated into tool use", "source": "https://api-docs.deepseek.com/news/news251201" }, { "name": "DeepSeek-V3.2-Speciale", "org": "DeepSeek", "country": "China", "date": "2025-12-01", "precision": "day", "category": "reasoning", "open_weights": true, "params": "685B (37B active)", "note": "Maxed-out reasoning variant with gold-level IMO and ICPC results", "source": "https://api-docs.deepseek.com/news/news251201" }, { "name": "HyperCLOVA X SEED 8B Omni", "org": "Naver", "country": "South Korea", "date": "2025-12-01", "precision": "month", "category": "multimodal", "open_weights": true, "params": "8B", "note": "Naver's open omni-modal model", "source": "https://huggingface.co/api/models?author=naver-hyperclovax&sort=createdAt&direction=-1&limit=100" }, { "name": "K2-V2", "org": "MBZUAI", "country": "UAE", "date": "2025-12-01", "precision": "month", "category": "language", "open_weights": true, "params": "70B", "note": "Fully open 70B dense model with released data and code", "source": "https://huggingface.co/IFM/K2-V2" }, { "name": "Runway Gen-4.5", "org": "Runway", "country": "USA", "date": "2025-12-01", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Ranked #1 on Artificial Analysis text-to-video leaderboard at launch", "source": "https://runwayml.com/research/introducing-runway-gen-4.5" }, { "name": "Trinity Mini", "org": "Arcee AI", "country": "USA", "date": "2025-12-01", "precision": "day", "category": "language", "open_weights": true, "params": "26B (3B active)", "note": "Open reasoning MoE from Arcee's US-built Trinity family", "source": "https://www.arcee.ai/blog/the-trinity-manifesto" }, { "name": "Trinity Nano Preview", "org": "Arcee AI", "country": "USA", "date": "2025-12-01", "precision": "day", "category": "language", "open_weights": true, "params": "6B", "note": "Experimental tiny MoE for low-resource chat", "source": "https://www.arcee.ai/blog/the-trinity-manifesto" }, { "name": "Vidu Q2 Image Generation", "org": "ShengShu", "country": "China", "date": "2025-12-01", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Vidu expands Q2 into image generation", "source": "https://www.prnewswire.com/news/shengshu-technology/" }, { "name": "Wan 2.6", "org": "Alibaba", "country": "China", "date": "2025-12-01", "precision": "month", "category": "video", "open_weights": false, "params": "", "note": "Multi-shot storytelling video with audio and reference-to-video", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Amazon Nova 2 Lite", "org": "Amazon", "country": "USA", "date": "2025-12-02", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Fast, cheap reasoning model with adjustable thinking intensity", "source": "https://aws.amazon.com/about-aws/whats-new/2025/12/nova-2-foundation-models-amazon-bedrock/" }, { "name": "Amazon Nova 2 Omni", "org": "Amazon", "country": "USA", "date": "2025-12-02", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Multimodal-input reasoning model that also generates images (preview)", "source": "https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-nova-2-omni-preview" }, { "name": "Amazon Nova 2 Pro (Preview)", "org": "Amazon", "country": "USA", "date": "2025-12-02", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Amazon's most intelligent model for complex multistep tasks", "source": "https://aws.amazon.com/about-aws/whats-new/2025/12/nova-2-foundation-models-amazon-bedrock/" }, { "name": "Amazon Nova 2 Sonic", "org": "Amazon", "country": "USA", "date": "2025-12-02", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-to-speech foundation model for real-time voice conversations", "source": "https://aws.amazon.com/blogs/aws/introducing-amazon-nova-2-sonic-next-generation-speech-to-speech-model-for-conversational-ai/" }, { "name": "GAIA-3", "org": "Wayve", "country": "UK", "date": "2025-12-02", "precision": "day", "category": "video", "open_weights": false, "params": "15B", "note": "Driving world model repurposed for safety evaluation", "source": "https://wayve.ai/thinking/gaia-3/" }, { "name": "GR-RL", "org": "ByteDance", "country": "China", "date": "2025-12-02", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "RL-trained robot policy; first real-robot shoe lacing", "source": "https://seed.bytedance.com/blog/gr-rl-a-breakthrough-in-dexterous-manipulation-achieving-the-first-real-robot-shoe-lacing-with-reinforcement-learning" }, { "name": "Kling O1", "org": "Kuaishou", "country": "China", "date": "2025-12-02", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Unified multimodal video generation and editing model", "source": "https://www.prnewswire.com/news/kuaishou-technology/" }, { "name": "Ministral 3", "org": "Mistral AI", "country": "France", "date": "2025-12-02", "precision": "day", "category": "language", "open_weights": true, "params": "14B", "note": "Open 3B/8B/14B edge models with instruct and reasoning variants", "source": "https://mistral.ai/news/mistral-3" }, { "name": "Mistral Large 3", "org": "Mistral AI", "country": "France", "date": "2025-12-02", "precision": "day", "category": "language", "open_weights": true, "params": "675B (41B active)", "note": "Mistral's open-weight frontier MoE under Apache 2.0", "source": "https://mistral.ai/news/mistral-3" }, { "name": "Hermes 4.3", "org": "Nous Research", "country": "USA", "date": "2025-12-03", "precision": "day", "category": "language", "open_weights": true, "params": "36B", "note": "Hermes 4.3 built on ByteDance Seed-OSS-36B", "source": "https://nousresearch.com/releases" }, { "name": "Seedream 4.5", "org": "ByteDance", "country": "China", "date": "2025-12-03", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Upgraded image generation and editing model", "source": "https://seed.bytedance.com/en/seedream4_5" }, { "name": "VibeVoice-Realtime-0.5B", "org": "Microsoft", "country": "USA", "date": "2025-12-03", "precision": "day", "category": "audio", "open_weights": true, "params": "0.5B", "note": "Real-time streaming text-to-speech with long-form generation", "source": "https://github.com/microsoft/VibeVoice" }, { "name": "HY 2.0 (Hunyuan 2.0)", "org": "Tencent", "country": "China", "date": "2025-12-05", "precision": "day", "category": "language", "open_weights": false, "params": "406B (32B active)", "note": "Next-gen MoE flagship in Think and Instruct versions", "source": "https://technode.com/2025/12/08/tencent-releases-hunyuan-2-0-its-next-generation-ai-model/" }, { "name": "Kling 2.6", "org": "Kuaishou", "country": "China", "date": "2025-12-05", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "First Kling with simultaneous audio-visual generation", "source": "https://www.prnewswire.com/news/kuaishou-technology/" }, { "name": "LongCat-Image", "org": "Meituan", "country": "China", "date": "2025-12-05", "precision": "day", "category": "image", "open_weights": true, "params": "6B", "note": "Compact open bilingual image generation and editing model", "source": "https://github.com/meituan-longcat/LongCat-Image" }, { "name": "Rnj-1", "org": "Essential AI", "country": "USA", "date": "2025-12-05", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Essential AI's first open base and instruct models", "source": "https://www.essential.ai/research/rnj-1" }, { "name": "GLM-4.6V", "org": "Zhipu AI", "country": "China", "date": "2025-12-08", "precision": "day", "category": "language", "open_weights": true, "params": "106B (12B active)", "note": "Open vision-language model with 128K context from Z.ai", "source": "https://docs.z.ai/release-notes/new-released" }, { "name": "Devstral 2", "org": "Mistral AI", "country": "France", "date": "2025-12-09", "precision": "day", "category": "code", "open_weights": true, "params": "123B", "note": "Open 123B agentic coding model launched with Mistral Vibe CLI", "source": "https://mistral.ai/news/devstral-2-vibe-cli" }, { "name": "Devstral Small 2", "org": "Mistral AI", "country": "France", "date": "2025-12-09", "precision": "day", "category": "code", "open_weights": true, "params": "24B", "note": "Apache-licensed local coding agent model", "source": "https://mistral.ai/news/devstral-2-vibe-cli" }, { "name": "Jais 2", "org": "G42", "country": "UAE", "date": "2025-12-09", "precision": "day", "category": "language", "open_weights": true, "params": "8B / 70B", "note": "Largest open Arabic-centric LLM trained from scratch", "source": "https://huggingface.co/papers/2608.13580" }, { "name": "Nomos 1", "org": "Nous Research", "country": "USA", "date": "2025-12-09", "precision": "day", "category": "science", "open_weights": true, "params": "30B", "note": "Open mathematical problem-solving and proof-writing model", "source": "https://nousresearch.com/releases" }, { "name": "GLM-ASR-2512", "org": "Zhipu AI", "country": "China", "date": "2025-12-10", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "ASR model claiming 0.0717 character error rate; smaller Nano variant open", "source": "https://docs.z.ai/release-notes/new-released" }, { "name": "AutoGLM-Phone-Multilingual", "org": "Zhipu AI", "country": "China", "date": "2025-12-11", "precision": "day", "category": "agent", "open_weights": true, "params": "9B", "note": "Open phone-control agent executing tasks across 50+ apps", "source": "https://docs.z.ai/release-notes/new-released" }, { "name": "GPT-5.2", "org": "OpenAI", "country": "USA", "date": "2025-12-11", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Shipped three weeks after Gemini 3 Pro amid OpenAI's 'code red'", "source": "https://en.wikipedia.org/wiki/GPT-5.2" }, { "name": "GPT-5.2 Pro", "org": "OpenAI", "country": "USA", "date": "2025-12-11", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Highest-compute GPT-5.2 tier for hardest problems", "source": "https://openai.com/index/introducing-gpt-5-2/" }, { "name": "GWM-1", "org": "Runway", "country": "USA", "date": "2025-12-11", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Runway's general world model built to simulate reality in real time", "source": "https://runwayml.com/research/introducing-runway-gwm-1" }, { "name": "Molmo 2", "org": "Allen Institute for AI", "country": "USA", "date": "2025-12-11", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Open video understanding, pointing and tracking model", "source": "https://allenai.org/blog/molmo2" }, { "name": "Rerank 4", "org": "Cohere", "country": "Canada", "date": "2025-12-11", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Cohere's most accurate reranker, Pro and Fast variants", "source": "https://cohere.com/blog/rerank-4" }, { "name": "SHARP", "org": "Apple", "country": "USA", "date": "2025-12-11", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Monocular 3D view synthesis in under a second", "source": "https://arxiv.org/abs/2512.10685" }, { "name": "Olmo 3.1", "org": "Allen Institute for AI", "country": "USA", "date": "2025-12-12", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32B", "note": "Extended RL run: Olmo 3.1 Think and Instruct 32B", "source": "https://allenai.org/blog/olmo3" }, { "name": "EuroLLM-22B", "org": "EuroLLM (Unbabel / IST)", "country": "Portugal", "date": "2025-12-14", "precision": "day", "category": "language", "open_weights": true, "params": "22B", "note": "Largest EuroLLM, 35 languages, trained on MareNostrum 5", "source": "https://huggingface.co/blog/eurollm-team/eurollm-22b" }, { "name": "Bolmo", "org": "Allen Institute for AI", "country": "USA", "date": "2025-12-15", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "First fully open byte-level LMs, byteified from Olmo 3", "source": "https://allenai.org/blog/bolmo" }, { "name": "Nemotron 3 Nano", "org": "NVIDIA", "country": "USA", "date": "2025-12-15", "precision": "day", "category": "language", "open_weights": true, "params": "30B (3B active)", "note": "Hybrid mixture-of-experts model opening the Nemotron 3 open family", "source": "https://nvidianews.nvidia.com/news/nvidia-debuts-nemotron-3-family-of-open-models" }, { "name": "Nemotron 3 Super", "org": "NVIDIA", "country": "USA", "date": "2025-12-15", "precision": "day", "category": "language", "open_weights": true, "params": "120B (12B active)", "note": "Announced Dec 2025; weights released March 11, 2026", "source": "https://developer.nvidia.com/blog/introducing-nemotron-3-super-an-open-hybrid-mamba-transformer-moe-for-agentic-reasoning/" }, { "name": "Nemotron 3 Ultra", "org": "NVIDIA", "country": "USA", "date": "2025-12-15", "precision": "day", "category": "language", "open_weights": true, "params": "550B (55B active)", "note": "Announced Dec 2025; weights released June 4, 2026", "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16" }, { "name": "FLUX.2 [max]", "org": "Black Forest Labs", "country": "Germany", "date": "2025-12-16", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Top-tier FLUX.2 variant", "source": "https://en.wikipedia.org/wiki/Flux_(text-to-image_model)" }, { "name": "gpt-image-1.5", "org": "OpenAI", "country": "USA", "date": "2025-12-16", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "OpenAI's latest and most advanced image generation model", "source": "https://platform.openai.com/docs/changelog" }, { "name": "Grok Voice Agent (speech-to-speech)", "org": "xAI", "country": "USA", "date": "2025-12-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-to-speech voice agent API on in-house audio models.", "source": "https://docs.x.ai/developers/release-notes" }, { "name": "Grok Voice Agent API", "org": "xAI", "country": "USA", "date": "2025-12-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-to-speech voice agent model offered via API", "source": "https://x.ai/news/grok-voice-agent-api" }, { "name": "LongCat-Video-Avatar", "org": "Meituan", "country": "China", "date": "2025-12-16", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open audio-driven character animation model", "source": "https://github.com/meituan-longcat/LongCat-Video" }, { "name": "MiMo-V2-Flash", "org": "Xiaomi", "country": "China", "date": "2025-12-16", "precision": "day", "category": "language", "open_weights": true, "params": "309B (15B active)", "note": "MIT-licensed MoE with hybrid sliding-window attention", "source": "https://finance.biggo.com/news/202512170422_Xiaomi_MiMo-V2-Flash_Open-Source_AI_Model_Launch" }, { "name": "Mistral Small Creative", "org": "Mistral AI", "country": "France", "date": "2025-12-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Labs model tuned for creative writing", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "SAM Audio", "org": "Meta", "country": "USA", "date": "2025-12-16", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Unified model separating sounds using multimodal prompts", "source": "https://ai.meta.com/blog/sam-audio/" }, { "name": "Seedance 1.5 Pro", "org": "ByteDance", "country": "China", "date": "2025-12-16", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Generates audio and video jointly in a single model", "source": "https://seed.bytedance.com/en/blog/sound-and-vision-all-in-one-take-the-official-release-of-seedance-1-5-pro" }, { "name": "TRELLIS.2", "org": "Microsoft", "country": "USA", "date": "2025-12-16", "precision": "day", "category": "image", "open_weights": true, "params": "4B", "note": "Native compact structured latents for 3D generation", "source": "https://arxiv.org/abs/2512.14692" }, { "name": "Gemini 3 Flash", "org": "Google DeepMind", "country": "USA", "date": "2025-12-17", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Fast frontier-class model rivaling larger models at lower cost", "source": "https://blog.google/products/gemini/gemini-3-flash/" }, { "name": "HY-World 1.5 (WorldPlay)", "org": "Tencent", "country": "China", "date": "2025-12-17", "precision": "day", "category": "video", "open_weights": true, "params": "8B", "note": "First open real-time interactive world model with long-term consistency", "source": "https://github.com/Tencent-Hunyuan/HY-WorldPlay" }, { "name": "Mistral OCR 3", "org": "Mistral AI", "country": "France", "date": "2025-12-17", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Third-generation document OCR model", "source": "https://mistral.ai/news" }, { "name": "FunctionGemma", "org": "Google DeepMind", "country": "USA", "date": "2025-12-18", "precision": "day", "category": "language", "open_weights": true, "params": "270M", "note": "Tiny Gemma specialized for function calling", "source": "https://blog.google/technology/developers/functiongemma/" }, { "name": "GPT-5.2-Codex", "org": "OpenAI", "country": "USA", "date": "2025-12-18", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "GPT-5.2 tuned for agentic coding, with stronger cybersecurity skills", "source": "https://simonwillison.net/2025/Dec/19/introducing-gpt-52-codex/" }, { "name": "HunyuanWorld 1.5 (WorldPlay)", "org": "Tencent", "country": "China", "date": "2025-12-18", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open real-time interactive world model for world creation and play", "source": "https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0" }, { "name": "HY-WorldPlay", "org": "Tencent", "country": "China", "date": "2025-12-18", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Real-time interactive world creation and play", "source": "https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0" }, { "name": "Luma Ray3 Modify", "org": "Luma AI", "country": "USA", "date": "2025-12-18", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Keyframe and character-reference controlled video modification", "source": "https://lumalabs.ai/news" }, { "name": "Seed1.8", "org": "ByteDance", "country": "China", "date": "2025-12-18", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Generalized agentic model for real-world tasks", "source": "https://seed.bytedance.com/blog/official-release-of-seed1-8-a-generalized-agentic-model" }, { "name": "T5Gemma 2", "org": "Google DeepMind", "country": "USA", "date": "2025-12-18", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Next generation of Gemma-based encoder-decoder models", "source": "https://blog.google/technology/developers/t5gemma-2/" }, { "name": "Cosmos-Reason2", "org": "NVIDIA", "country": "USA", "date": "2025-12-19", "precision": "day", "category": "robotics", "open_weights": true, "params": "8B", "note": "Open reasoning VLMs for physical-AI common sense and embodied reasoning", "source": "https://github.com/nvidia-cosmos/cosmos-reason2" }, { "name": "Kanana-2", "org": "Kakao", "country": "South Korea", "date": "2025-12-19", "precision": "day", "category": "language", "open_weights": true, "params": "30B (3B active)", "note": "MLA plus MoE agentic models including thinking variant", "source": "https://github.com/kakao/kanana-2" }, { "name": "Qwen-Image-Layered", "org": "Alibaba", "country": "China", "date": "2025-12-19", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Decomposes images into separately editable layers", "source": "https://github.com/QwenLM/Qwen-Image" }, { "name": "GLM-4.7", "org": "Zhipu AI", "country": "China", "date": "2025-12-22", "precision": "day", "category": "language", "open_weights": true, "params": "358B", "note": "Open coding and agent flagship upgrade from Z.ai", "source": "https://z.ai/blog/glm-4.7" }, { "name": "MiniMax-M2.1", "org": "MiniMax", "country": "China", "date": "2025-12-23", "precision": "day", "category": "language", "open_weights": true, "params": "229B", "note": "Stronger multi-language programming for real-world agents", "source": "https://www.minimax.io/news/minimax-m21" }, { "name": "Qwen-Image-Edit-2511", "org": "Alibaba", "country": "China", "date": "2025-12-23", "precision": "day", "category": "image", "open_weights": true, "params": "20B", "note": "Image editing update with multi-image support and better consistency", "source": "https://github.com/QwenLM/Qwen-Image" }, { "name": "Seed Prover 1.5", "org": "ByteDance", "country": "China", "date": "2025-12-24", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Agentic formal mathematical reasoning system", "source": "https://seed.bytedance.com/blog/seed-prover-1-5-advanced-mathematical-reasoning-through-a-novel-agentic-architecture" }, { "name": "HyperCLOVA X SEED 32B Think", "org": "NAVER", "country": "South Korea", "date": "2025-12-29", "precision": "day", "category": "reasoning", "open_weights": true, "params": "32B", "note": "NAVER's open 32B vision-language reasoning model", "source": "https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Think-32B" }, { "name": "A.X K1", "org": "SK Telecom", "country": "South Korea", "date": "2025-12-30", "precision": "day", "category": "language", "open_weights": true, "params": "519B (33B active)", "note": "SKT's 519B open MoE trained from scratch", "source": "https://huggingface.co/skt/A.X-K1" }, { "name": "HY-Motion 1.0", "org": "Tencent", "country": "China", "date": "2025-12-30", "precision": "day", "category": "video", "open_weights": true, "params": "1B", "note": "Billion-parameter text-to-3D-human-motion DiT", "source": "https://www.marktechpost.com/2025/12/31/tencent-released-tencent-hy-motion-1-0-a-billion-parameter-text-to-motion-model-built-on-the-diffusion-transformer-dit-architecture-and-flow-matching/" }, { "name": "HY-MT1.5", "org": "Tencent", "country": "China", "date": "2025-12-30", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Open 1.8B and 7B translation models succeeding Hunyuan-MT", "source": "https://github.com/Tencent-Hunyuan/Hunyuan-MT" }, { "name": "VAETKI", "org": "NC AI", "country": "South Korea", "date": "2025-12-30", "precision": "day", "category": "language", "open_weights": true, "params": "100B (10B active)", "note": "Open MoE from NC AI-led Korean consortium", "source": "https://huggingface.co/NC-AI-consortium-VAETKI/VAETKI" }, { "name": "K-EXAONE", "org": "LG", "country": "South Korea", "date": "2025-12-31", "precision": "day", "category": "language", "open_weights": true, "params": "236B (23B active)", "note": "LG's first large MoE, flagship for Korean sovereign AI", "source": "https://github.com/LG-AI-EXAONE/K-EXAONE" }, { "name": "Qwen-Image-2512", "org": "Alibaba", "country": "China", "date": "2025-12-31", "precision": "day", "category": "image", "open_weights": true, "params": "20B", "note": "Year-end Qwen-Image upgrade with more realistic humans and text", "source": "https://github.com/QwenLM/Qwen-Image" }, { "name": "Solar Open 100B", "org": "Upstage", "country": "South Korea", "date": "2025-12-31", "precision": "day", "category": "language", "open_weights": true, "params": "102B (12B active)", "note": "Upstage's flagship open MoE trained entirely from scratch", "source": "https://huggingface.co/upstage/Solar-Open-100B" }, { "name": "Inworld TTS-1.5", "org": "Inworld AI", "country": "USA", "date": "2026-01-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Low-latency proprietary text-to-speech model", "source": "https://thursdai.news/releases/2026-01" }, { "name": "LightOnOCR-2-1B", "org": "LightOn", "country": "France", "date": "2026-01-01", "precision": "month", "category": "vision", "open_weights": true, "params": "1B", "note": "Apache-2.0 end-to-end OCR model at under $0.01 per 1k pages", "source": "https://huggingface.co/lightonai/LightOnOCR-2-1B" }, { "name": "PersonaPlex-7B", "org": "NVIDIA", "country": "USA", "date": "2026-01-01", "precision": "month", "category": "audio", "open_weights": true, "params": "7B", "note": "Open full-duplex conversational speech model with persona control", "source": "https://huggingface.co/nvidia/personaplex-7b-v1" }, { "name": "NitroGen", "org": "NVIDIA", "country": "USA", "date": "2026-01-04", "precision": "day", "category": "games", "open_weights": true, "params": "500M", "note": "Foundation model for generalist game-playing agents", "source": "https://arxiv.org/abs/2601.02427" }, { "name": "Falcon-H1-Arabic", "org": "TII", "country": "UAE", "date": "2026-01-05", "precision": "day", "category": "language", "open_weights": true, "params": "3B / 7B / 34B", "note": "Hybrid-architecture Arabic model family", "source": "https://falconllm.tii.ae/abu-dhabi-tii-launches-falcon-h1-arabic-establishing-the-worlds-leading-arabic-ai-model.html" }, { "name": "Falcon-H1R-7B", "org": "TII", "country": "UAE", "date": "2026-01-05", "precision": "day", "category": "reasoning", "open_weights": true, "params": "7B", "note": "Hybrid Transformer-Mamba small reasoning model with 256K context", "source": "https://arxiv.org/abs/2601.02346" }, { "name": "LFM2.5", "org": "Liquid AI", "country": "USA", "date": "2026-01-05", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Next-generation on-device Liquid model family", "source": "https://www.liquid.ai/blog/introducing-lfm2-5-the-next-generation-of-on-device-ai" }, { "name": "Yuan3.0 Flash", "org": "IEIT Systems", "country": "China", "date": "2026-01-05", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Open multimodal enterprise LLM from Inspur's Yuan lab", "source": "https://arxiv.org/abs/2601.01718" }, { "name": "NousCoder-14B", "org": "Nous Research", "country": "USA", "date": "2026-01-06", "precision": "day", "category": "code", "open_weights": true, "params": "14B", "note": "Open competitive-programming code model", "source": "https://nousresearch.com/releases" }, { "name": "Jamba2", "org": "AI21 Labs", "country": "Israel", "date": "2026-01-08", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Hybrid SSM-Transformer enterprise model family under Apache 2.0", "source": "https://news.smol.ai/issues/26-01-08-not-much" }, { "name": "Qwen3-VL-Embedding", "org": "Alibaba", "country": "China", "date": "2026-01-08", "precision": "day", "category": "language", "open_weights": true, "params": "2B / 8B", "note": "Multimodal embedding plus reranker pair for text, image, video retrieval", "source": "https://news.smol.ai/issues/26-01-08-not-much" }, { "name": "Niji 7", "org": "Midjourney", "country": "USA", "date": "2026-01-09", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Anime-focused Midjourney model, first Niji update since 2024", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "Scribe v2", "org": "ElevenLabs", "country": "USA", "date": "2026-01-09", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Batch speech recognition across 90+ languages", "source": "https://elevenlabs.io/blog" }, { "name": "AgentCPM-Explore", "org": "OpenBMB", "country": "China", "date": "2026-01-12", "precision": "day", "category": "agent", "open_weights": true, "params": "4B", "note": "4B agent foundation model open-sourced with full training infrastructure", "source": "https://huggingface.co/openbmb/AgentCPM-Explore" }, { "name": "Baichuan-M3", "org": "Baichuan", "country": "China", "date": "2026-01-13", "precision": "day", "category": "language", "open_weights": true, "params": "235B", "note": "Open medical model claiming world-leading clinical performance", "source": "https://pandaily.com/baichuan-ai-open-sources-world-leading-medical-ai-model-m3" }, { "name": "MedASR", "org": "Google", "country": "USA", "date": "2026-01-13", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Medical speech-to-text model released alongside MedGemma 1.5", "source": "https://research.google/blog/next-generation-medical-image-interpretation-with-medgemma-15-and-medical-speech-to-text-with-medasr/" }, { "name": "MedGemma 1.5", "org": "Google", "country": "USA", "date": "2026-01-13", "precision": "day", "category": "science", "open_weights": true, "params": "4B", "note": "Open medical model adding CT, MRI and 3D imaging interpretation", "source": "https://research.google/blog/next-generation-medical-image-interpretation-with-medgemma-15-and-medical-speech-to-text-with-medasr/" }, { "name": "PixVerse R1", "org": "PixVerse", "country": "China", "date": "2026-01-13", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Marketed as a real-time interactive video world model", "source": "https://news.smol.ai/issues/26-01-13-not-much" }, { "name": "Pocket TTS", "org": "Kyutai", "country": "France", "date": "2026-01-13", "precision": "day", "category": "audio", "open_weights": true, "params": "100M", "note": "Laptop-CPU text-to-speech with voice cloning, no GPU needed", "source": "https://kyutai.org/blog/2026-01-13-pocket-tts" }, { "name": "GLM-Image", "org": "Zhipu AI", "country": "China", "date": "2026-01-14", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Hybrid autoregressive plus diffusion image model strong at text rendering", "source": "https://huggingface.co/zai-org/GLM-Image" }, { "name": "LongCat-Flash-Thinking-2601", "org": "Meituan", "country": "China", "date": "2026-01-14", "precision": "day", "category": "reasoning", "open_weights": true, "params": "560B (27B active)", "note": "Updated MIT-licensed MoE reasoning model from Meituan", "source": "https://huggingface.co/meituan-longcat/LongCat-Flash-Thinking-2601" }, { "name": "PersonaPlex", "org": "NVIDIA", "country": "USA", "date": "2026-01-14", "precision": "day", "category": "audio", "open_weights": true, "params": "7B", "note": "Full-duplex conversational speech with voice and role control", "source": "https://arxiv.org/abs/2602.06053" }, { "name": "Step-Audio R1.1", "org": "StepFun", "country": "China", "date": "2026-01-14", "precision": "day", "category": "audio", "open_weights": true, "params": "32B", "note": "Audio reasoning speech model leading Big Bench Audio at 96.4%", "source": "https://github.com/stepfun-ai/Step-Audio-R1" }, { "name": "Step3-VL-10B", "org": "StepFun", "country": "China", "date": "2026-01-14", "precision": "day", "category": "language", "open_weights": true, "params": "10B", "note": "Compact open vision-language reasoning model", "source": "https://arxiv.org/abs/2601.09668" }, { "name": "Falcon-H1-Tiny", "org": "TII", "country": "UAE", "date": "2026-01-15", "precision": "day", "category": "language", "open_weights": true, "params": "~90M", "note": "Sub-100M-parameter models for edge reasoning, coding and tool calls", "source": "https://news.smol.ai/issues/26-01-15-openresponses" }, { "name": "FLUX.2 [klein]", "org": "Black Forest Labs", "country": "Germany", "date": "2026-01-15", "precision": "day", "category": "image", "open_weights": true, "params": "4B / 9B", "note": "Sub-second small image generation and editing models; 4B Apache 2.0", "source": "https://news.smol.ai/issues/26-01-15-openresponses" }, { "name": "TranslateGemma", "org": "Google DeepMind", "country": "USA", "date": "2026-01-15", "precision": "day", "category": "language", "open_weights": true, "params": "4B / 12B / 27B", "note": "Open Gemma 3-based translation models covering 55 languages", "source": "https://blog.google/innovation-and-ai/technology/developers-tools/translategemma/" }, { "name": "Music-2.5", "org": "MiniMax", "country": "China", "date": "2026-01-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music model focused on fine-grained control and realism", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "GLM-4.7-Flash", "org": "Zhipu AI", "country": "China", "date": "2026-01-19", "precision": "day", "category": "language", "open_weights": true, "params": "30B (3B active)", "note": "Lightweight local coding and agent model, widely adopted locally", "source": "https://huggingface.co/zai-org/GLM-4.7-Flash" }, { "name": "LFM2.5-1.2B-Thinking", "org": "Liquid AI", "country": "USA", "date": "2026-01-20", "precision": "day", "category": "reasoning", "open_weights": true, "params": "1.2B", "note": "On-device reasoning model running in under 1GB memory", "source": "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Thinking" }, { "name": "Waypoint-1", "org": "Overworld", "country": "USA", "date": "2026-01-20", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Real-time interactive world model running on consumer GPUs", "source": "https://huggingface.co/Overworld/Waypoint-1-Small" }, { "name": "Yuan3.0 Ultra", "org": "IEIT Systems", "country": "China", "date": "2026-01-20", "precision": "day", "category": "language", "open_weights": true, "params": "1010B (68.8B active)", "note": "Trillion-parameter enterprise MoE; weights opened in early March", "source": "https://arxiv.org/abs/2601.14327" }, { "name": "HiRO-ACE", "org": "Allen Institute for AI", "country": "USA", "date": "2026-01-21", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Climate emulator plus 3 km precipitation downscaler on one GPU", "source": "https://allenai.org/blog/hiro-ace" }, { "name": "Rho-alpha", "org": "Microsoft", "country": "USA", "date": "2026-01-21", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Phi-derived VLA with tactile sensing for bimanual robots", "source": "https://www.microsoft.com/en-us/research/story/advancing-ai-for-the-physical-world/" }, { "name": "VibeVoice-ASR", "org": "Microsoft", "country": "USA", "date": "2026-01-21", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Transcribes 60-minute audio in one pass with speakers and timestamps", "source": "https://github.com/microsoft/VibeVoice" }, { "name": "D4RT", "org": "Google DeepMind", "country": "UK", "date": "2026-01-22", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Unified 4D scene reconstruction and tracking, up to 300x faster", "source": "https://deepmind.google/blog/d4rt-teaching-ai-to-see-the-world-in-four-dimensions/" }, { "name": "Qwen3-TTS", "org": "Alibaba", "country": "China", "date": "2026-01-22", "precision": "day", "category": "audio", "open_weights": true, "params": "0.6B / 1.7B", "note": "Open-sourced TTS family with voice design and cloning", "source": "https://github.com/QwenLM/Qwen3-TTS" }, { "name": "MiniMax Speech 2.8", "org": "MiniMax", "country": "China", "date": "2026-01-23", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Sound tags and 10-second voice cloning", "source": "https://www.minimax.io/news/minimax-speech-28" }, { "name": "Speech-2.8", "org": "MiniMax", "country": "China", "date": "2026-01-23", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "TTS with natural sound tags and lifelike voice", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "Earth-2 Medium Range (Atlas)", "org": "NVIDIA", "country": "USA", "date": "2026-01-26", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open global medium-range weather forecasting model in Earth-2 family", "source": "https://huggingface.co/nvidia/atlas-era5" }, { "name": "Earth-2 Nowcasting (StormScope)", "org": "NVIDIA", "country": "USA", "date": "2026-01-26", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open storm nowcasting model driven by satellite and radar data", "source": "https://huggingface.co/nvidia/stormscope-goes-mrms" }, { "name": "HunyuanImage 3.0-Instruct", "org": "Tencent", "country": "China", "date": "2026-01-26", "precision": "day", "category": "image", "open_weights": true, "params": "80B (13B active)", "note": "Reasoning-driven image editing and generation on an 80B MoE", "source": "https://github.com/Tencent-Hunyuan/HunyuanImage-3.0" }, { "name": "Luma Ray3.14", "org": "Luma AI", "country": "USA", "date": "2026-01-26", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Native 1080p; 3x cheaper and 4x faster than Ray3", "source": "https://lumalabs.ai/news" }, { "name": "Qwen3-Max-Thinking", "org": "Alibaba", "country": "China", "date": "2026-01-26", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Proprietary flagship reasoning model with adaptive tool use", "source": "https://qwen.ai/blog?id=qwen3-max-thinking" }, { "name": "Solar Pro 3", "org": "Upstage", "country": "South Korea", "date": "2026-01-26", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Doubles agentic performance with stronger reasoning", "source": "https://www.upstage.ai/blog" }, { "name": "C-RADIOv4", "org": "NVIDIA", "country": "USA", "date": "2026-01-27", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Agglomerative vision foundation backbone distilled from multiple teachers", "source": "https://huggingface.co/nvidia/C-RADIOv4-H" }, { "name": "DeepSeek-OCR 2", "org": "DeepSeek", "country": "China", "date": "2026-01-27", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "OCR with learned reading order via Visual Causal Flow encoder", "source": "https://huggingface.co/deepseek-ai/DeepSeek-OCR-2" }, { "name": "Helix 02", "org": "Figure AI", "country": "USA", "date": "2026-01-27", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Full-body autonomous humanoid control without teleoperation", "source": "https://www.figure.ai/news/helix-02" }, { "name": "K2 Think V2", "org": "MBZUAI", "country": "UAE", "date": "2026-01-27", "precision": "day", "category": "reasoning", "open_weights": true, "params": "70B", "note": "Fully open reasoning model built on K2-V2", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Kimi K2.5", "org": "Moonshot AI", "country": "China", "date": "2026-01-27", "precision": "day", "category": "language", "open_weights": true, "params": "1T (32B active)", "note": "Native image+video open model with 100-subagent swarm mode", "source": "https://news.smol.ai/issues/26-01-27-kimi-k25" }, { "name": "LingBot-VLA", "org": "Ant Group", "country": "China", "date": "2026-01-27", "precision": "day", "category": "robotics", "open_weights": true, "params": "4B", "note": "Open vision-language-action model scaling with real robot data", "source": "https://github.com/Robbyant/lingbot-vla" }, { "name": "Lucy 2", "org": "Decart", "country": "Israel", "date": "2026-01-27", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Real-time autoregressive video editing model with tech report", "source": "https://news.smol.ai/issues/26-01-27-kimi-k25" }, { "name": "SERA", "org": "Allen Institute for AI", "country": "USA", "date": "2026-01-27", "precision": "day", "category": "code", "open_weights": true, "params": "32B", "note": "Open coding agents; 54.2% SWE-Bench Verified, fine-tunable on private repos", "source": "https://allenai.org/blog/open-coding-agents" }, { "name": "Trinity Large", "org": "Arcee AI", "country": "USA", "date": "2026-01-27", "precision": "day", "category": "language", "open_weights": true, "params": "400B (13B active)", "note": "US open MoE pretrained from scratch; preview weights released", "source": "https://news.smol.ai/issues/26-01-27-kimi-k25" }, { "name": "Z-Image", "org": "Alibaba", "country": "China", "date": "2026-01-27", "precision": "day", "category": "image", "open_weights": true, "params": "6B", "note": "Non-distilled base model following Z-Image-Turbo", "source": "https://github.com/Tongyi-MAI/Z-Image" }, { "name": "MiniMax Music 2.5", "org": "MiniMax", "country": "China", "date": "2026-01-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Paragraph-level control and high-fidelity sound", "source": "https://www.minimax.io/news/minimax-music-25" }, { "name": "Mureka V8", "org": "Kunlun Tech", "country": "China", "date": "2026-01-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music model claiming global No. 1", "source": "https://pandaily.com/kunlun-tian-gong-unveils-mureka-v8-music-model-claims-global-no-1-and-connects-ai-music-to-commercial-distribution" }, { "name": "LingBot-World", "org": "Ant Group", "country": "China", "date": "2026-01-29", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open real-time interactive world model with technical report", "source": "https://github.com/Robbyant/lingbot-world" }, { "name": "MOVA", "org": "OpenMOSS", "country": "China", "date": "2026-01-29", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open foundation model for synchronized video and audio generation", "source": "https://github.com/OpenMOSS/MOVA" }, { "name": "PaddleOCR-VL-1.5", "org": "Baidu", "country": "China", "date": "2026-01-29", "precision": "day", "category": "vision", "open_weights": true, "params": "0.9B", "note": "Compact document-parsing VLM shipped with PaddleOCR 3.4", "source": "https://huggingface.co/PaddlePaddle/PaddleOCR-VL-1.5" }, { "name": "Project Genie (Genie 3)", "org": "Google DeepMind", "country": "USA", "date": "2026-01-29", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "First public access to Genie 3 interactive world generation", "source": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie/" }, { "name": "Qwen3-ASR", "org": "Alibaba", "country": "China", "date": "2026-01-29", "precision": "day", "category": "audio", "open_weights": true, "params": "0.6B / 1.7B", "note": "Open speech recognition for 52 languages plus forced aligner", "source": "https://github.com/QwenLM/Qwen3-ASR" }, { "name": "SkyReels-V3", "org": "Skywork AI", "country": "Singapore", "date": "2026-01-29", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open audio-to-video and reference-to-video generation models", "source": "https://github.com/SkyworkAI/SkyReels-V3" }, { "name": "Vidu Q3", "org": "Shengshu", "country": "China", "date": "2026-01-30", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "16-second synced audio-video; top-2 on AA video arena", "source": "https://news.qq.com/rain/a/20260130A06EDJ00" }, { "name": "ACE-Step 1.5", "org": "ACE Studio / StepFun", "country": "China", "date": "2026-01-31", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open music generator billed as open-source Suno alternative", "source": "https://arxiv.org/abs/2602.00744" }, { "name": "Eleven v3 Conversational", "org": "ElevenLabs", "country": "USA", "date": "2026-02-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Context-aware TTS model for ElevenAgents expressive mode", "source": "https://en.wikipedia.org/wiki/ElevenLabs" }, { "name": "Grok Imagine 1.0", "org": "xAI", "country": "USA", "date": "2026-02-01", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "10-second 720p video generation with improved audio", "source": "https://x.com/xai/status/2018164753810764061" }, { "name": "LLaDA2.1", "org": "Ant Group", "country": "China", "date": "2026-02-01", "precision": "month", "category": "language", "open_weights": true, "params": "", "note": "Diffusion language model accelerated via token editing", "source": "https://github.com/inclusionAI/LLaDA2.X" }, { "name": "MOSS-TTS", "org": "OpenMOSS", "country": "China", "date": "2026-02-01", "precision": "month", "category": "audio", "open_weights": true, "params": "", "note": "Open TTS family with realtime and dialogue variants", "source": "https://huggingface.co/OpenMOSS-Team/MOSS-TTS" }, { "name": "Seedream 5.0", "org": "ByteDance", "country": "China", "date": "2026-02-01", "precision": "month", "category": "image", "open_weights": false, "params": "", "note": "Next-generation Seedream image model, including a Lite tier", "source": "https://news.smol.ai/issues/26-02-11-glm-5" }, { "name": "SkyReels-V4", "org": "Kunlun Tech", "country": "China", "date": "2026-02-01", "precision": "month", "category": "video", "open_weights": false, "params": "", "note": "Ranked top-2 then No. 1 on Artificial Analysis video arena", "source": "https://www.pingwest.com/w/311695" }, { "name": "Qwen3-Coder-Next", "org": "Alibaba", "country": "China", "date": "2026-02-02", "precision": "day", "category": "code", "open_weights": true, "params": "80B (3B active)", "note": "Small hybrid-attention model for agentic coding", "source": "https://en.wikipedia.org/wiki/Qwen" }, { "name": "Sarvam Audio", "org": "Sarvam AI", "country": "India", "date": "2026-02-02", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech recognition model going beyond transcription", "source": "https://www.sarvam.ai/blogs" }, { "name": "Step 3.5 Flash", "org": "StepFun", "country": "China", "date": "2026-02-02", "precision": "day", "category": "language", "open_weights": true, "params": "196B (11B active)", "note": "Fast sparse MoE for agents with 256K context", "source": "https://static.stepfun.com/blog/step-3.5-flash/" }, { "name": "GLM-OCR", "org": "Zhipu AI", "country": "China", "date": "2026-02-03", "precision": "day", "category": "vision", "open_weights": true, "params": "0.9B", "note": "Sub-1B document OCR model ranked #1 on OmniDocBench v1.5", "source": "https://news.smol.ai/issues/26-02-03-not-much" }, { "name": "MiniCPM-o 4.5", "org": "OpenBMB", "country": "China", "date": "2026-02-03", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Full-duplex omni-modal live-streaming model for local devices", "source": "https://github.com/OpenBMB/MiniCPM-o" }, { "name": "Protenix-v1", "org": "ByteDance", "country": "China", "date": "2026-02-03", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open biomolecular structure predictor reaching AF3-level accuracy", "source": "https://seed.bytedance.com/en/protenix_pxdesign" }, { "name": "DreamZero", "org": "NVIDIA", "country": "USA", "date": "2026-02-04", "precision": "day", "category": "robotics", "open_weights": true, "params": "14B", "note": "World action model acting as zero-shot robot policy", "source": "https://arxiv.org/abs/2602.15922" }, { "name": "Intern-S1-Pro", "org": "Shanghai AI Lab", "country": "China", "date": "2026-02-04", "precision": "day", "category": "science", "open_weights": true, "params": "1T (22B active)", "note": "Trillion-parameter open scientific multimodal MoE", "source": "https://huggingface.co/internlm/Intern-S1-Pro" }, { "name": "Voxtral Realtime", "org": "Mistral AI", "country": "France", "date": "2026-02-04", "precision": "day", "category": "audio", "open_weights": true, "params": "4B", "note": "Apache-2.0 streaming transcription model with sub-500ms latency", "source": "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602" }, { "name": "Voxtral Transcribe 2", "org": "Mistral AI", "country": "France", "date": "2026-02-04", "precision": "day", "category": "audio", "open_weights": true, "params": "4B (Realtime)", "note": "Batch and realtime transcription; realtime 4B model open-weight", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "Bulbul V3", "org": "Sarvam AI", "country": "India", "date": "2026-02-05", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Expressive production text-to-speech for Indian languages", "source": "https://www.sarvam.ai/blogs" }, { "name": "Claude Opus 4.6", "org": "Anthropic", "country": "USA", "date": "2026-02-05", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "First Opus with 1M context; introduced agent teams", "source": "https://www.anthropic.com/news/claude-opus-4-6" }, { "name": "GPT-5.3-Codex", "org": "OpenAI", "country": "USA", "date": "2026-02-05", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Agentic coding model; OpenAI's first High cybersecurity-capability rating", "source": "https://openai.com/index/introducing-gpt-5-3-codex/" }, { "name": "Kling 3.0", "org": "Kuaishou", "country": "China", "date": "2026-02-05", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Native 4K, multi-shot sequencing and integrated audio", "source": "https://www.prnewswire.com/news-releases/kling-ai-launches-3-0-model-ushering-in-an-era-where-everyone-can-be-a-director-302679944.html" }, { "name": "Kling Image 3.0", "org": "Kuaishou", "country": "China", "date": "2026-02-05", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Cinematic storytelling image model in Kling 3.0 family", "source": "https://www.prnewswire.com/news/kling-ai/" }, { "name": "Kling Video 3.0 Omni", "org": "Kuaishou", "country": "China", "date": "2026-02-05", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Unified multimodal member of the Kling 3.0 family", "source": "https://www.prnewswire.com/news/kling-ai/" }, { "name": "Sarvam Vision", "org": "Sarvam AI", "country": "India", "date": "2026-02-05", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Document-intelligence vision model for Indian languages", "source": "https://www.sarvam.ai/blogs" }, { "name": "DreamDojo", "org": "NVIDIA", "country": "USA", "date": "2026-02-06", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Generalist robot world model trained on large-scale human video", "source": "https://arxiv.org/abs/2602.06949" }, { "name": "LongCat-Flash-Lite", "org": "Meituan", "country": "China", "date": "2026-02-06", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Lightweight open MoE strong at agents and code", "source": "http://news.17173.com/content/02062026/180535539.shtml" }, { "name": "Waymo World Model", "org": "Waymo", "country": "USA", "date": "2026-02-06", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Generative driving simulator built on DeepMind's Genie 3", "source": "https://news.smol.ai/issues/26-02-06-not-much" }, { "name": "Composer 1.5", "org": "Cursor", "country": "USA", "date": "2026-02-09", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Cursor's in-house agentic coding model update", "source": "https://cursor.com/blog/composer-1-5" }, { "name": "Seedream 5.0 Lite", "org": "ByteDance", "country": "China", "date": "2026-02-09", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Reasoning image model with real-time web search", "source": "https://eu.36kr.com/en/p/3677025348395653" }, { "name": "Aletheia", "org": "Google DeepMind", "country": "USA", "date": "2026-02-10", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Deep Think-based agent autonomously solving research-level math problems", "source": "https://arxiv.org/abs/2602.10177" }, { "name": "FireRed-Image-Edit 1.0", "org": "Xiaohongshu", "country": "China", "date": "2026-02-10", "precision": "day", "category": "image", "open_weights": true, "params": "", "note": "Open image-editing model claiming state-of-the-art benchmark results", "source": "https://github.com/FireRedTeam/FireRed-Image-Edit" }, { "name": "IsoDDE (Isomorphic Labs Drug Design Engine)", "org": "Isomorphic Labs", "country": "UK", "date": "2026-02-10", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Drug-design system more than doubling AlphaFold 3 on hard ligand cases", "source": "https://www.isomorphiclabs.com/articles/the-isomorphic-labs-drug-design-engine-unlocks-a-new-frontier" }, { "name": "Ming-Flash-Omni 2.0", "org": "Ant Group", "country": "China", "date": "2026-02-10", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Open full-modality understanding-and-generation model", "source": "https://pandaily.com/ant-group-open-sources-full-modality-model-ming-flash-omni-2-0" }, { "name": "Qwen-Image-2.0", "org": "Alibaba", "country": "China", "date": "2026-02-10", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Unified generation and editing with native 2K and typography", "source": "https://github.com/QwenLM/Qwen-Image" }, { "name": "Saaras V3", "org": "Sarvam AI", "country": "India", "date": "2026-02-10", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech recognition and translation model for Indian languages", "source": "https://www.sarvam.ai/blogs" }, { "name": "GLM-5", "org": "Zhipu AI", "country": "China", "date": "2026-02-11", "precision": "day", "category": "language", "open_weights": true, "params": "744B (40B active)", "note": "MIT-licensed agentic engineering model; #1 open model on Text Arena", "source": "https://z.ai/blog/glm-5" }, { "name": "MiniCPM-SALA", "org": "OpenBMB", "country": "China", "date": "2026-02-11", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Hybrid sparse-linear attention for million-token context", "source": "https://huggingface.co/openbmb/MiniCPM-SALA" }, { "name": "Spark X2", "org": "iFlytek", "country": "China", "date": "2026-02-11", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Flagship benchmarked against top international models, domestic compute", "source": "https://wap.eastmoney.com/a/202602113648088954.html" }, { "name": "GPT-5.3-Codex-Spark", "org": "OpenAI", "country": "USA", "date": "2026-02-12", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Over 1,000 tokens/s coding model served on Cerebras", "source": "https://openai.com/index/introducing-gpt-5-3-codex-spark/" }, { "name": "Hibiki-Zero", "org": "Kyutai", "country": "France", "date": "2026-02-12", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Simultaneous speech translation trained without aligned data", "source": "https://kyutai.org/blog" }, { "name": "MiniMax-M2.5", "org": "MiniMax", "country": "China", "date": "2026-02-12", "precision": "day", "category": "language", "open_weights": true, "params": "230B (10B active)", "note": "Agent-native coding model at 80.2% SWE-bench Verified", "source": "https://www.minimax.io/news/minimax-m25" }, { "name": "Seedance 2.0", "org": "ByteDance", "country": "China", "date": "2026-02-12", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Viral photoreal clips; drew Hollywood cease-and-desist letters", "source": "https://en.wikipedia.org/wiki/Seedance" }, { "name": "Ring-2.5-1T", "org": "Ant Group", "country": "China", "date": "2026-02-13", "precision": "day", "category": "reasoning", "open_weights": true, "params": "1T", "note": "Upgraded trillion-parameter thinking model", "source": "https://quantumzeitgeist.com/ant-group-thinking-models-ai-benchmarks/" }, { "name": "JoyAI-LLM-Flash", "org": "JD", "country": "China", "date": "2026-02-14", "precision": "day", "category": "language", "open_weights": true, "params": "48B (3B active)", "note": "JD's first open-weight LLM", "source": "https://finance.sina.cn/tech/2026-02-15/detail-inhmwwmm8181938.d.html" }, { "name": "Seed 2.0", "org": "ByteDance", "country": "China", "date": "2026-02-14", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Doubao-Seed-2.0 Pro/Lite/Mini; Pro pitched against GPT-5.2", "source": "https://seed.bytedance.com/blog/seed-2-0-official-launch" }, { "name": "BitDance", "org": "ByteDance", "country": "China", "date": "2026-02-15", "precision": "day", "category": "image", "open_weights": true, "params": "14B", "note": "Autoregressive image generator predicting binary visual tokens", "source": "https://arxiv.org/abs/2602.14041" }, { "name": "Ling-2.5-1T", "org": "Ant Group", "country": "China", "date": "2026-02-16", "precision": "day", "category": "language", "open_weights": true, "params": "1T", "note": "Upgraded trillion-parameter instruct model", "source": "https://www.fintechweekly.com/news/ant-group-ling-2-5-1t-ring-2-5-1t-open-source-ai-models" }, { "name": "Qwen3.5", "org": "Alibaba", "country": "China", "date": "2026-02-16", "precision": "day", "category": "language", "open_weights": true, "params": "397B (17B active)", "note": "Agent-focused open flagship; hosted as Qwen3.5-Plus", "source": "https://github.com/QwenLM/Qwen3.5" }, { "name": "Qwen3.5-397B-A17B", "org": "Alibaba", "country": "China", "date": "2026-02-16", "precision": "day", "category": "language", "open_weights": true, "params": "397B (17B active)", "note": "First open Qwen3.5; native multimodal hybrid-attention MoE", "source": "https://qwen.ai/blog?id=qwen3.5" }, { "name": "Claude Sonnet 4.6", "org": "Anthropic", "country": "USA", "date": "2026-02-17", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "New claude.ai default with 1M-token context beta", "source": "https://www.anthropic.com/news/claude-sonnet-4-6" }, { "name": "Grok 4.20", "org": "xAI", "country": "USA", "date": "2026-02-17", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Public-beta flagship with multi-agent mode and 2M context", "source": "https://en.wikipedia.org/wiki/Grok_(chatbot)" }, { "name": "Param-2", "org": "BharatGen", "country": "India", "date": "2026-02-17", "precision": "day", "category": "language", "open_weights": true, "params": "17B", "note": "Larger BharatGen sovereign model under research license", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Param2 17B", "org": "BharatGen", "country": "India", "date": "2026-02-17", "precision": "day", "category": "language", "open_weights": true, "params": "17B (2.4B active)", "note": "Government-backed Indian sovereign MoE with thinking variant", "source": "https://huggingface.co/bharatgenai/Param2-17B-A2.4B-Thinking" }, { "name": "Tiny Aya", "org": "Cohere", "country": "Canada", "date": "2026-02-17", "precision": "day", "category": "language", "open_weights": true, "params": "3.35B", "note": "Phone-sized multilingual model covering 70+ languages", "source": "https://news.smol.ai/issues/26-02-17-sonnet-46" }, { "name": "jina-embeddings-v5-text", "org": "Jina AI", "country": "Germany", "date": "2026-02-18", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Fifth-generation multilingual text embeddings in small and nano sizes", "source": "https://huggingface.co/jinaai/jina-embeddings-v5-text-small" }, { "name": "Lyria 3", "org": "Google DeepMind", "country": "USA", "date": "2026-02-18", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music model with lyrics and vocals shipped in Gemini app", "source": "https://blog.google/innovation-and-ai/products/gemini-app/lyria-3/" }, { "name": "Sarvam-105B", "org": "Sarvam AI", "country": "India", "date": "2026-02-18", "precision": "day", "category": "language", "open_weights": true, "params": "105B (~9B active)", "note": "India's largest homegrown MoE; powers the Indus chat app", "source": "https://en.wikipedia.org/wiki/Sarvam_AI" }, { "name": "Sarvam-30B", "org": "Sarvam AI", "country": "India", "date": "2026-02-18", "precision": "day", "category": "language", "open_weights": true, "params": "30B", "note": "Smaller Sarvam MoE for Indian languages", "source": "https://en.wikipedia.org/wiki/Sarvam_AI" }, { "name": "ZUNA", "org": "Zyphra", "country": "USA", "date": "2026-02-18", "precision": "day", "category": "science", "open_weights": true, "params": "380M", "note": "Open EEG foundation model for brain-computer interfaces", "source": "https://huggingface.co/Zyphra/ZUNA" }, { "name": "Gemini 3.1 Pro", "org": "Google DeepMind", "country": "USA", "date": "2026-02-19", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "77.1% ARC-AGI-2, more than double Gemini 3 Pro", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/" }, { "name": "Tri-21B-Think", "org": "Trillion Labs", "country": "South Korea", "date": "2026-02-19", "precision": "day", "category": "reasoning", "open_weights": true, "params": "21B", "note": "Korean Apache-2.0 reasoning model released as preview", "source": "https://huggingface.co/trillionlabs/Tri-21B-Think" }, { "name": "GPT-Audio-1.5", "org": "OpenAI", "country": "USA", "date": "2026-02-23", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Upgraded audio in/out model for Chat Completions", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "gpt-realtime-1.5", "org": "OpenAI", "country": "USA", "date": "2026-02-23", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-to-speech voice-agent model with better tool calling", "source": "https://developers.openai.com/api/docs/models/gpt-realtime-1.5" }, { "name": "LFM2-24B-A2B", "org": "Liquid AI", "country": "USA", "date": "2026-02-24", "precision": "day", "category": "language", "open_weights": true, "params": "24B (2.3B active)", "note": "Largest LFM2 MoE, fits laptops within 32GB", "source": "https://huggingface.co/LiquidAI/LFM2-24B-A2B" }, { "name": "Mercury 2", "org": "Inception", "country": "USA", "date": "2026-02-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Reasoning diffusion LLM at roughly 1,000 output tokens/s", "source": "https://news.smol.ai/issues/26-02-24-claude-code" }, { "name": "Qwen3.5-122B-A10B", "org": "Alibaba", "country": "China", "date": "2026-02-24", "precision": "day", "category": "language", "open_weights": true, "params": "122B (10B active)", "note": "Medium Qwen3.5 MoE supporting 1M+ context", "source": "https://github.com/QwenLM/Qwen3.5" }, { "name": "Qwen3.5-27B", "org": "Alibaba", "country": "China", "date": "2026-02-24", "precision": "day", "category": "language", "open_weights": true, "params": "27B", "note": "Dense medium-size Qwen3.5 model", "source": "https://github.com/QwenLM/Qwen3.5" }, { "name": "Qwen3.5-35B-A3B", "org": "Alibaba", "country": "China", "date": "2026-02-24", "precision": "day", "category": "language", "open_weights": true, "params": "35B (3B active)", "note": "Local-agent favorite running on 32GB consumer hardware", "source": "https://github.com/QwenLM/Qwen3.5" }, { "name": "Qwen3.5-Flash", "org": "Alibaba", "country": "China", "date": "2026-02-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Hosted API version of Qwen3.5-35B-A3B", "source": "https://news.smol.ai/issues/26-02-24-claude-code" }, { "name": "Nano Banana 2 (Gemini 3.1 Flash Image)", "org": "Google DeepMind", "country": "USA", "date": "2026-02-26", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Pro-level image quality at Flash speed; new Gemini default", "source": "https://blog.google/innovation-and-ai/technology/ai/nano-banana-2/" }, { "name": "pplx-embed", "org": "Perplexity", "country": "USA", "date": "2026-02-26", "precision": "day", "category": "language", "open_weights": true, "params": "0.6B / 4B", "note": "MIT-licensed web-scale retrieval embeddings with context variant", "source": "https://news.smol.ai/issues/26-02-26-nanobanana2" }, { "name": "Chandra OCR 2", "org": "Datalab", "country": "USA", "date": "2026-03-01", "precision": "month", "category": "vision", "open_weights": true, "params": "4B", "note": "OCR model at 85.9% olmOCR-bench across 90+ languages", "source": "https://news.smol.ai/issues/26-03-19-not-much" }, { "name": "Falcon Perception", "org": "TII", "country": "UAE", "date": "2026-03-01", "precision": "month", "category": "vision", "open_weights": true, "params": "0.6B", "note": "Early-fusion VLM for open-vocabulary grounding and segmentation", "source": "https://huggingface.co/tiiuae/Falcon-Perception" }, { "name": "GigaChat 3.1 Ultra", "org": "Sber", "country": "Russia", "date": "2026-03-01", "precision": "month", "category": "language", "open_weights": true, "params": "702B (36B active)", "note": "Updated MIT-licensed Sber flagship MoE", "source": "https://huggingface.co/ai-sage/GigaChat3.1-702B-A36B" }, { "name": "GLM-5-Turbo", "org": "Zhipu AI", "country": "China", "date": "2026-03-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Faster GLM-5 variant for GLM Coding Plan agents", "source": "https://news.smol.ai/issues/26-03-26-not-much" }, { "name": "LTX-2.3", "org": "Lightricks", "country": "Israel", "date": "2026-03-01", "precision": "month", "category": "video", "open_weights": true, "params": "", "note": "Open video model shipped with a local desktop editor", "source": "https://en.wikipedia.org/wiki/LTX_(world_model)" }, { "name": "Qwen3.5-Max-Preview", "org": "Alibaba", "country": "China", "date": "2026-03-01", "precision": "month", "category": "language", "open_weights": false, "params": "", "note": "Proprietary flagship preview; #3 in Arena math", "source": "https://news.smol.ai/issues/26-03-19-not-much" }, { "name": "SWE-1.6", "org": "Cognition", "country": "USA", "date": "2026-03-01", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Preview of Cognition's fast in-house software engineering model", "source": "https://cognition.ai/blog/swe-1-6-preview" }, { "name": "Qwen3.5 Small (0.8B/2B/4B/9B)", "org": "Alibaba", "country": "China", "date": "2026-03-02", "precision": "day", "category": "language", "open_weights": true, "params": "0.8B-9B", "note": "Native multimodal edge models; 9B rivals much larger models", "source": "https://github.com/QwenLM/Qwen3.5" }, { "name": "Qwen3.5-9B", "org": "Alibaba", "country": "China", "date": "2026-03-02", "precision": "day", "category": "language", "open_weights": true, "params": "9B", "note": "Small Qwen3.5 series: 9B, 4B, 2B, 0.8B", "source": "https://github.com/QwenLM/Qwen3.5" }, { "name": "Gemini 3.1 Flash-Lite", "org": "Google DeepMind", "country": "USA", "date": "2026-03-03", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Fastest, cheapest Gemini 3 model at $0.25 per million input tokens", "source": "https://siliconangle.com/2026/03/03/google-launches-speedy-gemini-3-1-flash-lite-model-preview/" }, { "name": "GPT-5.3 Instant", "org": "OpenAI", "country": "USA", "date": "2026-03-03", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "New ChatGPT default with fewer caveats and refusals", "source": "https://9to5mac.com/2026/03/03/openai-releases-gpt-5-3-instant-update-to-make-chatgpt-less-cringe/" }, { "name": "Helios", "org": "Peking University", "country": "China", "date": "2026-03-04", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Real-time long video generation on one GPU, with ByteDance and Canva", "source": "https://github.com/PKU-YuanGroup/Helios" }, { "name": "Phi-4-reasoning-vision", "org": "Microsoft", "country": "USA", "date": "2026-03-04", "precision": "day", "category": "reasoning", "open_weights": true, "params": "15B", "note": "Multimodal reasoning Phi with SigLIP-2 vision encoder", "source": "https://huggingface.co/microsoft/Phi-4-reasoning-vision-15B" }, { "name": "Phi-4-reasoning-vision-15B", "org": "Microsoft", "country": "USA", "date": "2026-03-04", "precision": "day", "category": "reasoning", "open_weights": true, "params": "15B", "note": "Compact open multimodal reasoning model", "source": "https://huggingface.co/microsoft/Phi-4-reasoning-vision-15B" }, { "name": "GPT-5.4", "org": "OpenAI", "country": "USA", "date": "2026-03-05", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Thinking model with 1M context and native computer use", "source": "https://openai.com/index/introducing-gpt-5-4/" }, { "name": "GPT-5.4 Pro", "org": "OpenAI", "country": "USA", "date": "2026-03-05", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Highest-compute tier; set new FrontierMath records", "source": "https://openai.com/index/introducing-gpt-5-4/" }, { "name": "KARL", "org": "Databricks", "country": "USA", "date": "2026-03-05", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "RL-trained knowledge agent for grounded enterprise document reasoning", "source": "https://news.smol.ai/issues/26-03-05-gpt54" }, { "name": "Olmo Hybrid", "org": "Allen Institute for AI", "country": "USA", "date": "2026-03-05", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Fully open 7B mixing attention with linear-RNN layers", "source": "https://news.smol.ai/issues/26-03-05-gpt54" }, { "name": "Granite 4.0 1B Speech", "org": "IBM", "country": "USA", "date": "2026-03-06", "precision": "day", "category": "audio", "open_weights": true, "params": "1B", "note": "Compact open speech recognition model in Granite 4.0 family", "source": "https://huggingface.co/ibm-granite/granite-4.0-1b-speech" }, { "name": "Gemini Embedding 2", "org": "Google DeepMind", "country": "USA", "date": "2026-03-10", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "First natively multimodal embedding model: text, image, video, audio, PDF", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-embedding-2/" }, { "name": "Grok 4.20 Multi-Agent", "org": "xAI", "country": "USA", "date": "2026-03-10", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "API variant orchestrating multiple collaborating agents for deep research.", "source": "https://docs.x.ai/developers/release-notes" }, { "name": "TADA", "org": "Hume AI", "country": "USA", "date": "2026-03-10", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open TTS claiming zero content hallucinations and 5x speed", "source": "https://news.smol.ai/issues/26-03-10-ami-labs" }, { "name": "MolmoBot", "org": "Allen Institute for AI", "country": "USA", "date": "2026-03-11", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Open manipulation models trained purely on synthetic MolmoSpaces data", "source": "https://allenai.org/blog/molmobot" }, { "name": "Qianfan-OCR", "org": "Baidu", "country": "China", "date": "2026-03-11", "precision": "day", "category": "vision", "open_weights": true, "params": "4B", "note": "End-to-end document intelligence model unifying OCR tasks", "source": "https://arxiv.org/abs/2603.13398" }, { "name": "Reka Edge", "org": "Reka AI", "country": "USA", "date": "2026-03-11", "precision": "day", "category": "vision", "open_weights": true, "params": "7B", "note": "Low-latency VLM for physical AI and edge deployment", "source": "https://news.smol.ai/issues/26-03-11-not-much" }, { "name": "Mistral Moderation 2", "org": "Mistral AI", "country": "France", "date": "2026-03-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Moderation model with 128k context and jailbreak detection", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "Grok Text to Speech", "org": "xAI", "country": "USA", "date": "2026-03-16", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Standalone expressive text-to-speech API made generally available.", "source": "https://docs.x.ai/developers/release-notes" }, { "name": "Kimodo", "org": "NVIDIA", "country": "USA", "date": "2026-03-16", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Controllable human and humanoid motion generation from 700h mocap", "source": "https://huggingface.co/nvidia/Kimodo-SOMA-RP-v1" }, { "name": "Leanstral", "org": "Mistral AI", "country": "France", "date": "2026-03-16", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Open Lean 4 proof-engineering model for trustworthy code", "source": "https://mistral.ai/news" }, { "name": "Mamba-3", "org": "Carnegie Mellon University", "country": "USA", "date": "2026-03-16", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "State-space architecture for inference-heavy workloads, with Princeton and Cartesia", "source": "https://arxiv.org/abs/2603.15569" }, { "name": "Mistral Small 4", "org": "Mistral AI", "country": "France", "date": "2026-03-16", "precision": "day", "category": "language", "open_weights": true, "params": "119B (6B active)", "note": "Apache-2.0 multimodal MoE with hybrid reasoning modes", "source": "https://mistral.ai/news/mistral-small-4/" }, { "name": "Nemotron 3 Nano 4B", "org": "NVIDIA", "country": "USA", "date": "2026-03-16", "precision": "day", "category": "language", "open_weights": true, "params": "4B", "note": "Smallest Nemotron 3 hybrid model for edge devices", "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16" }, { "name": "V-JEPA 2.1", "org": "Meta", "country": "USA", "date": "2026-03-16", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Self-supervised video model learning dense, temporally consistent features", "source": "https://github.com/facebookresearch/vjepa2" }, { "name": "GPT-5.4 mini", "org": "OpenAI", "country": "USA", "date": "2026-03-17", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Over 2x faster than GPT-5 mini; available to free ChatGPT users", "source": "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/" }, { "name": "GPT-5.4 nano", "org": "OpenAI", "country": "USA", "date": "2026-03-17", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "API-only smallest GPT-5.4 variant aimed at subagents", "source": "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/" }, { "name": "Holotron-12B", "org": "H Company", "country": "France", "date": "2026-03-17", "precision": "day", "category": "agent", "open_weights": true, "params": "12B", "note": "Open hybrid-SSM computer-use model built with NVIDIA", "source": "https://news.smol.ai/issues/26-03-17-not-much" }, { "name": "Midjourney V8", "org": "Midjourney", "country": "USA", "date": "2026-03-17", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Alpha of first new major Midjourney model since V7", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "Rakuten AI 3.0", "org": "Rakuten", "country": "Japan", "date": "2026-03-17", "precision": "day", "category": "language", "open_weights": true, "params": "671B (37B active)", "note": "~700B Japanese-optimized MoE released under Apache-2.0", "source": "https://huggingface.co/Rakuten/RakutenAI-3.0" }, { "name": "MiMo-V2-Omni", "org": "Xiaomi", "country": "China", "date": "2026-03-18", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Omni-modal model across vision, audio and video", "source": "https://en.wikipedia.org/wiki/Xiaomi_MiMo" }, { "name": "MiMo-V2-Pro", "org": "Xiaomi", "country": "China", "date": "2026-03-18", "precision": "day", "category": "language", "open_weights": false, "params": "1T", "note": "Xiaomi's API-only trillion-parameter agentic flagship", "source": "https://en.wikipedia.org/wiki/Xiaomi_MiMo" }, { "name": "MiMo-V2-TTS", "org": "Xiaomi", "country": "China", "date": "2026-03-18", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Expressive TTS with singing and Chinese dialects", "source": "https://en.wikipedia.org/wiki/Xiaomi_MiMo" }, { "name": "MiniMax-M2.7", "org": "MiniMax", "country": "China", "date": "2026-03-18", "precision": "day", "category": "language", "open_weights": true, "params": "230B", "note": "Self-evolving agent model scoring 56.2% on SWE-Pro", "source": "https://www.minimax.io/news/minimax-m27-en" }, { "name": "MolmoPoint", "org": "Allen Institute for AI", "country": "USA", "date": "2026-03-18", "precision": "day", "category": "vision", "open_weights": true, "params": "8B", "note": "Pointing VLMs for images, video, and GUIs", "source": "https://allenai.org/blog/molmopoint" }, { "name": "Alpamayo 1.5", "org": "NVIDIA", "country": "USA", "date": "2026-03-19", "precision": "day", "category": "robotics", "open_weights": true, "params": "10B", "note": "Steerable update of NVIDIA's reasoning driving VLA", "source": "https://huggingface.co/nvidia/Alpamayo-1.5-10B" }, { "name": "Composer 2", "org": "Cursor", "country": "USA", "date": "2026-03-19", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Frontier coding model continued-pretrained from Kimi K2.5", "source": "https://news.smol.ai/issues/26-03-19-not-much" }, { "name": "MAI-Image-2", "org": "Microsoft", "country": "USA", "date": "2026-03-19", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Debuted #5 on Image Arena with better text rendering", "source": "https://news.smol.ai/issues/26-03-19-not-much" }, { "name": "Nemotron-Cascade 2", "org": "NVIDIA", "country": "USA", "date": "2026-03-19", "precision": "day", "category": "reasoning", "open_weights": true, "params": "30B (3B active)", "note": "Cascade-RL reasoning MoE claiming olympiad gold-level math", "source": "https://huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B" }, { "name": "Namazu", "org": "Sakana AI", "country": "Japan", "date": "2026-03-23", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Japan-tuned alpha model family powering Sakana Chat", "source": "https://news.smol.ai/issues/26-03-23-not-much" }, { "name": "Uni-1", "org": "Luma AI", "country": "USA", "date": "2026-03-23", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Model that reasons and generates pixels simultaneously", "source": "https://news.smol.ai/issues/26-03-23-not-much" }, { "name": "Voxtral TTS", "org": "Mistral AI", "country": "France", "date": "2026-03-23", "precision": "day", "category": "audio", "open_weights": true, "params": "4B", "note": "Text-to-speech with zero-shot voice cloning, multilingual", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "MolmoWeb", "org": "Allen Institute for AI", "country": "USA", "date": "2026-03-24", "precision": "day", "category": "agent", "open_weights": true, "params": "4B / 8B", "note": "Open browser agent claiming open-weight SOTA on web benchmarks", "source": "https://news.smol.ai/issues/26-03-24-not-much" }, { "name": "HY-OmniWeaving", "org": "Tencent", "country": "China", "date": "2026-03-25", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Unified video generation model from Tencent Hunyuan", "source": "https://huggingface.co/tencent/HY-OmniWeaving" }, { "name": "LongCat-Next", "org": "Meituan", "country": "China", "date": "2026-03-25", "precision": "day", "category": "multimodal", "open_weights": true, "params": "68.5B (3B active)", "note": "Discrete-token native multimodal MoE for text, vision, audio", "source": "https://huggingface.co/meituan-longcat/LongCat-Next" }, { "name": "Lyria 3 Pro", "org": "Google DeepMind", "country": "USA", "date": "2026-03-25", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Three-minute songs with intro/verse/chorus structure control", "source": "https://techcrunch.com/2026/03/25/google-launches-lyria-3-pro-music-generation-model/" }, { "name": "Cohere Transcribe", "org": "Cohere", "country": "Canada", "date": "2026-03-26", "precision": "day", "category": "audio", "open_weights": true, "params": "2B", "note": "Apache-2.0 ASR at 5.42% WER on Open ASR Leaderboard", "source": "https://huggingface.co/CohereLabs/cohere-transcribe-03-2026" }, { "name": "Gemini 3.1 Flash Live", "org": "Google DeepMind", "country": "USA", "date": "2026-03-26", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Real-time voice and vision model powering Gemini Live", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-flash-live/" }, { "name": "Suno v5.5", "org": "Suno", "country": "USA", "date": "2026-03-26", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Adds own-voice singing and personalized custom models", "source": "https://en.wikipedia.org/wiki/Suno_AI" }, { "name": "TRIBE v2", "org": "Meta", "country": "USA", "date": "2026-03-26", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Trimodal brain encoder predicting fMRI responses zero-shot", "source": "https://news.smol.ai/issues/26-03-26-not-much" }, { "name": "Granite 4.0 3B Vision", "org": "IBM", "country": "USA", "date": "2026-03-27", "precision": "day", "category": "vision", "open_weights": true, "params": "3B", "note": "Compact open vision-language model in Granite 4.0 family", "source": "https://huggingface.co/ibm-granite/granite-4.0-3b-vision" }, { "name": "Matrix-Game 3.0", "org": "Skywork AI", "country": "Singapore", "date": "2026-03-27", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Real-time streaming interactive world model with long-horizon memory", "source": "https://github.com/SkyworkAI/Matrix-Game" }, { "name": "SAM 3.1", "org": "Meta", "country": "USA", "date": "2026-03-27", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Object multiplexing roughly doubles video tracking throughput", "source": "https://github.com/facebookresearch/sam3" }, { "name": "harrier-oss-v1", "org": "Microsoft", "country": "USA", "date": "2026-03-30", "precision": "day", "category": "language", "open_weights": true, "params": "270M / 0.6B / 27B", "note": "Open embedding family claiming SOTA on multilingual MTEB v2", "source": "https://huggingface.co/microsoft/harrier-oss-v1-27b" }, { "name": "PixVerse V6", "org": "PixVerse", "country": "China", "date": "2026-03-30", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Multi-shot video with audio for creative and agentic workflows", "source": "https://www.prnewswire.com/news-releases/pixverse-launches-v6-advancing-ai-video-generation-across-creative-and-agentic-workflows-302728386.html" }, { "name": "Qwen3.5-Omni", "org": "Alibaba", "country": "China", "date": "2026-03-30", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Native omni model with audio-visual vibe coding demo", "source": "https://news.smol.ai/issues/26-03-30-not-much" }, { "name": "Holo3", "org": "H Company", "country": "France", "date": "2026-03-31", "precision": "day", "category": "agent", "open_weights": true, "params": "35B (3B active)", "note": "Computer-use model scoring 77.8% on OSWorld-Verified", "source": "https://www.hcompany.ai/holo3" }, { "name": "LFM2.5-350M", "org": "Liquid AI", "country": "USA", "date": "2026-03-31", "precision": "day", "category": "language", "open_weights": true, "params": "350M", "note": "Tiny LFM2.5 model for constrained devices", "source": "https://www.liquid.ai/blog/lfm2-5-350m-no-size-left-behind" }, { "name": "Veo 3.1 Lite", "org": "Google DeepMind", "country": "UK", "date": "2026-03-31", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Most cost-efficient Veo video generation model", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "DeepL voice-to-voice translation", "org": "DeepL", "country": "Germany", "date": "2026-04-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Voice-to-voice translation in more than 40 languages", "source": "https://en.wikipedia.org/wiki/DeepL_Translator" }, { "name": "Music-2.6", "org": "MiniMax", "country": "China", "date": "2026-04-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Music model update improving covers and bass", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "OpenAI Privacy Filter", "org": "OpenAI", "country": "USA", "date": "2026-04-01", "precision": "month", "category": "language", "open_weights": true, "params": "1.5B (50M active)", "note": "Apache 2.0 PII-masking model; rare OpenAI open-weight release", "source": "https://huggingface.co/openai/privacy-filter" }, { "name": "Qwen3.6-Plus", "org": "Alibaba", "country": "China", "date": "2026-04-01", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "API flagship with sharply improved agentic coding", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Trinity-Large-Thinking", "org": "Arcee AI", "country": "USA", "date": "2026-04-01", "precision": "day", "category": "reasoning", "open_weights": true, "params": "400B (13B active)", "note": "Reasoning version of Trinity Large tuned for agents", "source": "https://www.arcee.ai/blog/trinity-large-thinking" }, { "name": "Wan2.7-Image", "org": "Alibaba", "country": "China", "date": "2026-04-01", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Image generation model of the Wan2.7 generation", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Gemma 4", "org": "Google DeepMind", "country": "UK", "date": "2026-04-02", "precision": "day", "category": "language", "open_weights": true, "params": "E2B, E4B, 26B A4B, 31B", "note": "Apache 2.0 open family with vision/audio input, up to 256K context", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "GEN-1", "org": "Generalist AI", "country": "USA", "date": "2026-04-02", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Embodied model claiming mastery of simple physical tasks", "source": "https://generalistai.com/blog/gen-1" }, { "name": "GLM-5V-Turbo", "org": "Zhipu AI", "country": "China", "date": "2026-04-02", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Vision-language model in the GLM-5 generation", "source": "https://github.com/zai-org/GLM-V" }, { "name": "MAI-Transcribe-1", "org": "Microsoft", "country": "USA", "date": "2026-04-02", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Microsoft AI in-house speech recognition model covering 25 languages", "source": "https://microsoft.ai/news/state-of-the-art-speech-recognition-with-mai-transcribe-1/" }, { "name": "Wan2.7", "org": "Alibaba", "country": "China", "date": "2026-04-03", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Video generation adding editing of visuals, plot and dialogue", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Claude Mythos Preview", "org": "Anthropic", "country": "USA", "date": "2026-04-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Withheld from public over cyber risk; Project Glasswing partners only", "source": "https://en.wikipedia.org/wiki/Claude_Mythos" }, { "name": "GLM-5.1", "org": "Zhipu AI", "country": "China", "date": "2026-04-07", "precision": "day", "category": "language", "open_weights": true, "params": "744B", "note": "MIT-licensed open flagship for long-horizon agentic engineering", "source": "https://en.wikipedia.org/wiki/Zhipu_AI" }, { "name": "WildDet3D", "org": "Allen Institute for AI", "country": "USA", "date": "2026-04-07", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Open monocular 3D object detection in the wild", "source": "https://allenai.org/blog/wilddet3d" }, { "name": "LFM2.5-VL-450M", "org": "Liquid AI", "country": "USA", "date": "2026-04-08", "precision": "day", "category": "language", "open_weights": true, "params": "450M", "note": "Tiny structured-vision VLM for edge to cloud", "source": "https://www.liquid.ai/blog/lfm2-5-vl-450m" }, { "name": "Muse Spark", "org": "Meta", "country": "USA", "date": "2026-04-08", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "First Meta Superintelligence Labs model; proprietary multimodal reasoning", "source": "https://en.wikipedia.org/wiki/Muse_Spark" }, { "name": "EXAONE 4.5", "org": "LG", "country": "South Korea", "date": "2026-04-09", "precision": "day", "category": "language", "open_weights": true, "params": "33B", "note": "LG's first open-weight vision-language model", "source": "https://github.com/LG-AI-EXAONE/EXAONE-4.5" }, { "name": "HY-Embodied-0.5", "org": "Tencent", "country": "China", "date": "2026-04-09", "precision": "day", "category": "robotics", "open_weights": true, "params": "2B", "note": "Open embodied foundation model for robot perception and planning", "source": "https://github.com/Tencent-Hunyuan/HY-Embodied" }, { "name": "Seeduplex", "org": "ByteDance", "country": "China", "date": "2026-04-09", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Full-duplex voice model integrated into Doubao app", "source": "https://pandaily.com/byte-dance-unveils-full-duplex-voice-model-seeduplex" }, { "name": "Gemini Robotics-ER 1.6", "org": "Google DeepMind", "country": "UK", "date": "2026-04-14", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Embodied-reasoning robotics model adding instrument reading", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "GPT-5.4-Cyber", "org": "OpenAI", "country": "USA", "date": "2026-04-14", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Cybersecurity-specialized variant of GPT-5.4", "source": "https://epoch.ai/data/notable_ai_models.csv" }, { "name": "Lyra 2.0", "org": "NVIDIA", "country": "USA", "date": "2026-04-14", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Generates explorable 3D worlds from a single image", "source": "https://huggingface.co/nvidia/Lyra-2.0" }, { "name": "MAI-Image-2-Efficient", "org": "Microsoft", "country": "USA", "date": "2026-04-14", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Faster, 41% cheaper variant of MAI-Image-2", "source": "https://microsoft.ai/news/mai-image-2-efficient/" }, { "name": "Midjourney V8.1", "org": "Midjourney", "country": "USA", "date": "2026-04-14", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Alpha update to the V8 image model, four weeks after V8", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "Qwen3.6-Max-Preview", "org": "Alibaba", "country": "China", "date": "2026-04-14", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Preview of the largest Qwen3.6 model", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "ERNIE-Image", "org": "Baidu", "country": "China", "date": "2026-04-15", "precision": "day", "category": "image", "open_weights": true, "params": "8B", "note": "Baidu's first open-weight text-to-image DiT, strong text rendering", "source": "https://cntechpost.com/2026/04/15/baidu-open-sources-ernie-image-model-bringing-top-tier-rendering-consumer-gpus/" }, { "name": "Gemini 3.1 Flash TTS", "org": "Google DeepMind", "country": "UK", "date": "2026-04-15", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Steerable expressive TTS with audio tags, 70+ languages", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "Qwen3.6-35B-A3B", "org": "Alibaba", "country": "China", "date": "2026-04-15", "precision": "day", "category": "language", "open_weights": true, "params": "35B (3B active)", "note": "First open Qwen3.6 weights; Apache 2.0 multimodal MoE", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Claude Opus 4.7", "org": "Anthropic", "country": "USA", "date": "2026-04-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "1M context, xhigh effort level, sharper vision, new tokenizer", "source": "https://platform.claude.com/docs/en/release-notes/overview" }, { "name": "GPT-Rosalind", "org": "OpenAI", "country": "USA", "date": "2026-04-16", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Life-sciences model for drug discovery; trusted-access research preview", "source": "https://openai.com/index/introducing-gpt-rosalind/" }, { "name": "HY-World 2.0", "org": "Tencent", "country": "China", "date": "2026-04-16", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open 3D world model producing editable 3D worlds, not just video", "source": "https://kr-asia.com/tencents-hy-world-2-0-moves-ai-beyond-video-into-editable-3d-worlds" }, { "name": "Qwen3.6-Flash", "org": "Alibaba", "country": "China", "date": "2026-04-16", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Low-cost fast tier of the Qwen3.6 API family", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "π0.7", "org": "Physical Intelligence", "country": "USA", "date": "2026-04-16", "precision": "day", "category": "robotics", "open_weights": false, "params": "5B", "note": "Next Physical Intelligence generalist VLA", "source": "https://arxiv.org/abs/2604.15483" }, { "name": "GR00T N1.7", "org": "NVIDIA", "country": "USA", "date": "2026-04-17", "precision": "day", "category": "robotics", "open_weights": true, "params": "3B", "note": "Trained on 20k hours of human egocentric video; dexterity scaling law", "source": "https://huggingface.co/blog/nvidia/gr00t-n1-7" }, { "name": "Grok 4.3", "org": "xAI", "country": "USA", "date": "2026-04-17", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "New pretrained model; native video input, generates slides, PDFs, spreadsheets.", "source": "https://www.roborhythms.com/grok-4-3-release-april-2026/" }, { "name": "Kimi K2.6", "org": "Moonshot AI", "country": "China", "date": "2026-04-20", "precision": "day", "category": "language", "open_weights": true, "params": "1T (32B active)", "note": "Open MoE for long-horizon coding with 300 sub-agent swarms", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Deep Research Max", "org": "Google DeepMind", "country": "UK", "date": "2026-04-21", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Most comprehensive Gemini research agent; MCP and collaborative planning", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "gpt-image-2", "org": "OpenAI", "country": "USA", "date": "2026-04-21", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "ChatGPT Images 2.0; first OpenAI image model with thinking", "source": "https://hidekazu-konishi.com/entry/openai_gpt_model_release_timeline.html" }, { "name": "HappyHorse-1.0", "org": "Alibaba", "country": "China", "date": "2026-04-22", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Video model that first topped Artificial Analysis arena pseudonymously", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Ling-2.6-flash", "org": "Ant Group", "country": "China", "date": "2026-04-22", "precision": "day", "category": "language", "open_weights": true, "params": "104B", "note": "Token-efficient MoE", "source": "https://www.businesswire.com/news/home/20260422256825/en/Ant-Group-Unveils-Ling-2.6-Flash-A-Major-Leap-in-AI-Efficiency" }, { "name": "MiMo-V2.5", "org": "Xiaomi", "country": "China", "date": "2026-04-22", "precision": "day", "category": "multimodal", "open_weights": true, "params": "310B (15B active)", "note": "Natively omnimodal MoE with 1M-token context", "source": "https://en.wikipedia.org/wiki/Xiaomi_MiMo" }, { "name": "MiMo-V2.5-Pro", "org": "Xiaomi", "country": "China", "date": "2026-04-22", "precision": "day", "category": "language", "open_weights": true, "params": "1.02T (42B active)", "note": "Xiaomi's trillion-parameter agentic model, MIT-licensed weights", "source": "https://en.wikipedia.org/wiki/Xiaomi_MiMo" }, { "name": "Qwen-Image-2.0-Pro", "org": "Alibaba", "country": "China", "date": "2026-04-22", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Pro tier of the Qwen-Image 2.0 generation and editing model", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Qwen3.6-27B", "org": "Alibaba", "country": "China", "date": "2026-04-22", "precision": "day", "category": "language", "open_weights": true, "params": "27B", "note": "Dense open model beating Qwen3.5-397B on coding benchmarks", "source": "https://news.smol.ai/issues/26-04-22-not-much" }, { "name": "GPT-5.5", "org": "OpenAI", "country": "USA", "date": "2026-04-23", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Codename Spud; flagship for agentic coding and computer use", "source": "https://en.wikipedia.org/wiki/GPT-5.5" }, { "name": "GPT-5.5 Pro", "org": "OpenAI", "country": "USA", "date": "2026-04-23", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Higher-compute GPT-5.5 variant for the most demanding work", "source": "https://hidekazu-konishi.com/entry/openai_gpt_model_release_timeline.html" }, { "name": "Grok Voice Think Fast 1.0", "org": "xAI", "country": "USA", "date": "2026-04-23", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Flagship speech-to-speech voice agent model for multi-step tool-calling workflows.", "source": "https://x.ai/news/grok-voice-think-fast-1" }, { "name": "Hy3 preview", "org": "Tencent", "country": "China", "date": "2026-04-23", "precision": "day", "category": "language", "open_weights": true, "params": "295B (21B active)", "note": "Tencent Hy team's open MoE for reasoning, coding and agents", "source": "https://huggingface.co/tencent/Hy3-preview" }, { "name": "Sapiens2", "org": "Meta", "country": "USA", "date": "2026-04-23", "precision": "day", "category": "vision", "open_weights": true, "params": "0.1B-5B", "note": "Human-centric ViTs pretrained on 1B human images", "source": "https://arxiv.org/abs/2604.21681" }, { "name": "Seed3D 2.0", "org": "ByteDance", "country": "China", "date": "2026-04-23", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Higher-precision 3D foundation model", "source": "https://seed.bytedance.com/en/blog/seed3d-2-0-released-higher-precision-and-greater-usability" }, { "name": "DeepSeek-V4-Flash", "org": "DeepSeek", "country": "China", "date": "2026-04-24", "precision": "day", "category": "language", "open_weights": true, "params": "284B (13B active)", "note": "Cheap V4 tier at $0.14/$0.28 per million tokens", "source": "https://api-docs.deepseek.com/updates" }, { "name": "DeepSeek-V4-Pro", "org": "DeepSeek", "country": "China", "date": "2026-04-24", "precision": "day", "category": "language", "open_weights": true, "params": "1.6T (49B active)", "note": "MIT-licensed 1M-context MoE; hybrid attention slashes KV cache", "source": "https://api-docs.deepseek.com/updates" }, { "name": "Sakana Fugu", "org": "Sakana AI", "country": "Japan", "date": "2026-04-24", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Multi-agent orchestration system served as a single model API", "source": "https://sakana.ai/fugu-release/" }, { "name": "Holotron3", "org": "H Company", "country": "France", "date": "2026-04-28", "precision": "day", "category": "agent", "open_weights": true, "params": "", "note": "Research model for agentic computer use", "source": "https://www.hcompany.ai/blog" }, { "name": "Laguna M.1", "org": "Poolside", "country": "USA", "date": "2026-04-28", "precision": "day", "category": "code", "open_weights": true, "params": "225B (23B active)", "note": "Poolside's larger in-house-trained agentic coding model", "source": "https://poolside.ai/blog" }, { "name": "Laguna XS.2", "org": "Poolside", "country": "USA", "date": "2026-04-28", "precision": "day", "category": "code", "open_weights": true, "params": "33B (3B active)", "note": "Apache 2.0 coding MoE that runs on a single GPU", "source": "https://poolside.ai/blog" }, { "name": "Ling-2.6-1T", "org": "Ant Group", "country": "China", "date": "2026-04-28", "precision": "day", "category": "language", "open_weights": true, "params": "1T", "note": "Trillion-parameter fast-thinking open model, MIT license", "source": "https://news.smol.ai/issues/26-04-29-not-much" }, { "name": "Mistral Medium 3.5", "org": "Mistral AI", "country": "France", "date": "2026-04-28", "precision": "day", "category": "language", "open_weights": true, "params": "128B", "note": "First merged flagship, replacing Magistral, Pixtral and Devstral", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Nemotron 3 Nano Omni", "org": "NVIDIA", "country": "USA", "date": "2026-04-28", "precision": "day", "category": "multimodal", "open_weights": true, "params": "31B (3B active)", "note": "Open omni model understanding video, audio, images and text", "source": "https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16" }, { "name": "SenseNova U1", "org": "SenseTime", "country": "China", "date": "2026-04-28", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Open VAE-free unified understanding-and-generation model", "source": "https://pandaily.com/sense-time-launches-sense-nova-u1-moving-toward-a-unified-model-era-of-understanding-and-generation" }, { "name": "Granite 4.1", "org": "IBM", "country": "USA", "date": "2026-04-29", "precision": "day", "category": "language", "open_weights": true, "params": "3B / 8B / 30B", "note": "Apache 2.0 dense enterprise models plus vision, speech, guardian", "source": "https://huggingface.co/ibm-granite/granite-4.1-8b" }, { "name": "Spark X2-Flash", "org": "iFlytek", "country": "China", "date": "2026-04-29", "precision": "day", "category": "language", "open_weights": false, "params": "30B", "note": "Low-cost 30B model with 256K context", "source": "https://www.pingwest.com/w/313404" }, { "name": "AI co-clinician", "org": "Google DeepMind", "country": "UK", "date": "2026-04-30", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Research medical agent system operating under physician authority", "source": "https://deepmind.google/blog/ai-co-clinician/" }, { "name": "ERNIE 5.1", "org": "Baidu", "country": "China", "date": "2026-04-30", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Preview Apr 30, official May 9; ~6% of peers' pretraining cost", "source": "https://x.com/Baidu_Inc/status/2049682555809788282" }, { "name": "Carbon", "org": "Hugging Face", "country": "USA", "date": "2026-05-01", "precision": "month", "category": "science", "open_weights": true, "params": "3B (family 500M-8B)", "note": "Open generative DNA foundation models with Zhongguancun Academy, TIGEM", "source": "https://huggingface.co/HuggingFaceBio/Carbon-3B" }, { "name": "ESMFold2", "org": "Biohub", "country": "USA", "date": "2026-05-01", "precision": "month", "category": "science", "open_weights": true, "params": "7B", "note": "All-atom structure prediction for proteins, DNA, RNA and ligands", "source": "https://huggingface.co/biohub/ESMFold2" }, { "name": "HiDream-O1-Image", "org": "HiDream", "country": "China", "date": "2026-05-01", "precision": "month", "category": "image", "open_weights": true, "params": "", "note": "HiDream's next-generation open image model", "source": "https://huggingface.co/HiDream-ai/HiDream-O1-Image" }, { "name": "Runway Aleph 2.0", "org": "Runway", "country": "USA", "date": "2026-05-01", "precision": "month", "category": "video", "open_weights": false, "params": "", "note": "Edit one frame to update a whole 30-second video", "source": "https://runway.com/product/aleph-2" }, { "name": "Toto 2.0", "org": "Datadog", "country": "USA", "date": "2026-05-01", "precision": "month", "category": "science", "open_weights": true, "params": "4M-2.5B", "note": "Open time-series forecasting foundation models, Apache 2.0", "source": "https://news.smol.ai/issues/26-05-14-not-much" }, { "name": "GPT-5.5 Instant", "org": "OpenAI", "country": "USA", "date": "2026-05-05", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "New ChatGPT default for all users, fewer hallucinations", "source": "https://hidekazu-konishi.com/entry/openai_gpt_model_release_timeline.html" }, { "name": "Inworld Realtime TTS-2", "org": "Inworld AI", "country": "USA", "date": "2026-05-05", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Closed-loop voice model adapting to how users talk", "source": "https://www.marktechpost.com/2026/05/05/inworld-ai-launches-realtime-tts-2-a-closed-loop-voice-model-that-adapts-to-how-you-actually-talk/" }, { "name": "Luma Uni-1.1", "org": "Luma AI", "country": "USA", "date": "2026-05-05", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Unified understanding-and-generation image model via API", "source": "https://lumalabs.ai/news" }, { "name": "MolmoAct 2", "org": "Allen Institute for AI", "country": "USA", "date": "2026-05-05", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Open robotics foundation model with Think and embodied-reasoning variants", "source": "https://allenai.org/blog/molmoact2" }, { "name": "SubQ", "org": "Subquadratic", "country": "USA", "date": "2026-05-05", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Sparse-attention model claiming multi-million (up to 12M) token context", "source": "https://subq.ai/" }, { "name": "Grok Imagine Quality Mode", "org": "xAI", "country": "USA", "date": "2026-05-06", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Higher-realism image generation with stronger text rendering via API.", "source": "https://x.ai/news/grok-imagine-quality-mode" }, { "name": "ZAYA1-8B", "org": "Zyphra", "country": "USA", "date": "2026-05-06", "precision": "day", "category": "reasoning", "open_weights": true, "params": "8.4B (760M active)", "note": "Reasoning MoE trained entirely on AMD MI300 GPUs", "source": "https://www.marktechpost.com/2026/05/06/zyphra-releases-zaya1-8b-a-reasoning-moe-trained-on-amd-hardware-that-punches-far-above-its-weight-class/" }, { "name": "GPT-5.5-Cyber", "org": "OpenAI", "country": "USA", "date": "2026-05-07", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Limited-preview cyber variant for vetted security teams", "source": "https://en.wikipedia.org/wiki/GPT-5.5" }, { "name": "GPT-Realtime-2", "org": "OpenAI", "country": "USA", "date": "2026-05-07", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-to-speech voice-agent model with GPT-5-class reasoning", "source": "https://news.smol.ai/issues/26-05-07-gpt-realtime-2" }, { "name": "GPT-Realtime-Translate", "org": "OpenAI", "country": "USA", "date": "2026-05-07", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Live speech translation from 70+ input languages", "source": "https://news.smol.ai/issues/26-05-07-gpt-realtime-2" }, { "name": "GPT-Realtime-Whisper", "org": "OpenAI", "country": "USA", "date": "2026-05-07", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Low-latency streaming speech-to-text model", "source": "https://news.smol.ai/issues/26-05-07-gpt-realtime-2" }, { "name": "ZAYA1-74B-Preview", "org": "Zyphra", "country": "USA", "date": "2026-05-07", "precision": "day", "category": "language", "open_weights": true, "params": "74B", "note": "Largest ZAYA1 model, preview release", "source": "https://www.zyphra.com/our-work" }, { "name": "EMO", "org": "Allen Institute for AI", "country": "USA", "date": "2026-05-08", "precision": "day", "category": "language", "open_weights": true, "params": "14B (1B active)", "note": "128-expert MoE studying expert clustering, with matched baseline", "source": "https://allenai.org/blog/emo" }, { "name": "SenseNova 6.7 Flash-Lite", "org": "SenseTime", "country": "China", "date": "2026-05-08", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Lightweight multimodal agent model cutting token use 60%", "source": "https://finance.sina.com.cn/stock/t/2026-05-08/doc-inhxefhi2265914.shtml" }, { "name": "ZAYA1-VL-8B", "org": "Zyphra", "country": "USA", "date": "2026-05-08", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Vision-language variant of ZAYA1", "source": "https://www.zyphra.com/our-work" }, { "name": "MiniCPM-V 4.6", "org": "ModelBest", "country": "China", "date": "2026-05-11", "precision": "day", "category": "language", "open_weights": true, "params": "1.3B", "note": "Ultra-efficient edge vision model", "source": "https://github.com/OpenBMB/MiniCPM-V" }, { "name": "TML-Interaction-Small", "org": "Thinking Machines Lab", "country": "USA", "date": "2026-05-11", "precision": "day", "category": "multimodal", "open_weights": false, "params": "276B (12B active)", "note": "Full-duplex interaction model for real-time audio/video collaboration", "source": "https://thinkingmachines.ai/blog/" }, { "name": "jina-embeddings-v5-omni", "org": "Jina AI", "country": "Germany", "date": "2026-05-12", "precision": "day", "category": "language", "open_weights": true, "params": "0.95B / 1.57B", "note": "Universal embeddings across text, images, audio and video", "source": "https://news.smol.ai/issues/26-05-12-not-much" }, { "name": "Perceptron Mk1", "org": "Perceptron", "country": "USA", "date": "2026-05-12", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Video and embodied reasoning model with spatial outputs", "source": "https://www.perceptron.inc/blog" }, { "name": "Ring-2.6-1T", "org": "Ant Group", "country": "China", "date": "2026-05-14", "precision": "day", "category": "reasoning", "open_weights": true, "params": "1T", "note": "Open thinking model claimed to beat GPT-5.4 on some tasks", "source": "https://gigazine.net/gsc_news/en/20260515-ring-2-6-1t-ai-china/" }, { "name": "ZAYA1-8B-Diffusion-Preview", "org": "Zyphra", "country": "USA", "date": "2026-05-14", "precision": "day", "category": "language", "open_weights": true, "params": "8B", "note": "Diffusion-based language model variant of ZAYA1", "source": "https://www.zyphra.com/our-work" }, { "name": "Intern-S2-Preview", "org": "Shanghai AI Lab", "country": "China", "date": "2026-05-15", "precision": "day", "category": "science", "open_weights": true, "params": "35B", "note": "35B scientific model rivaling trillion-scale S1-Pro via task scaling", "source": "https://startupfortune.com/internlm-is-making-scientific-ai-smaller-with-intern-s2-preview/" }, { "name": "Supertonic v3", "org": "Supertone", "country": "South Korea", "date": "2026-05-15", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "On-device TTS covering 31 languages with expression tags", "source": "https://www.marktechpost.com/2026/05/15/supertone-releases-supertonic-v3-on-device-text-to-speech-model-with-31-language-support-fewer-reading-failures-and-expression-tags/" }, { "name": "SANA-WM", "org": "NVIDIA", "country": "USA", "date": "2026-05-16", "precision": "day", "category": "video", "open_weights": true, "params": "2.6B", "note": "Open world model generating minute-long 720p video on one GPU", "source": "https://www.marktechpost.com/2026/05/16/nvidia-introduces-sana-wm-a-2-6b-parameter-open-source-world-model-that-generates-minute-scale-720p-video-on-a-single-gpu/" }, { "name": "Composer 2.5", "org": "Cursor", "country": "USA", "date": "2026-05-18", "precision": "day", "category": "code", "open_weights": false, "params": "1T", "note": "Cursor's in-IDE coding model built on Kimi K2.5", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Gemini 3.5 Flash", "org": "Google DeepMind", "country": "UK", "date": "2026-05-19", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "New default Gemini; beats 3.1 Pro on agentic coding", "source": "https://ai.google.dev/gemini-api/docs/changelog" }, { "name": "Gemini 3.5 Pro", "org": "Google DeepMind", "country": "UK", "date": "2026-05-19", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Announced at I/O while in internal use; rollout promised later", "source": "https://blog.google/innovation-and-ai/technology/ai/google-io-2026-all-our-announcements/" }, { "name": "Gemini Omni Flash", "org": "Google DeepMind", "country": "UK", "date": "2026-05-19", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "First Gemini Omni model: any-input generation, starting with video", "source": "https://blog.google/innovation-and-ai/technology/ai/google-io-2026-all-our-announcements/" }, { "name": "Grok Build 0.1", "org": "xAI", "country": "USA", "date": "2026-05-19", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Fast agentic coding model powering xAI's Grok Build CLI.", "source": "https://docs.x.ai/developers/release-notes" }, { "name": "MiniCPM5-1B", "org": "ModelBest", "country": "China", "date": "2026-05-19", "precision": "day", "category": "language", "open_weights": true, "params": "1B", "note": "First MiniCPM5 on-device model", "source": "https://github.com/OpenBMB/MiniCPM" }, { "name": "OlmoEarth v1.1", "org": "Allen Institute for AI", "country": "USA", "date": "2026-05-19", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Earth-observation models with up to 3x lower compute", "source": "https://allenai.org/blog/olmoearth-v1-1" }, { "name": "Qwen3.5-LiveTranslate-Flash", "org": "Alibaba", "country": "China", "date": "2026-05-19", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Real-time audio-video interpretation from 60 input languages", "source": "https://www.marktechpost.com/2026/05/20/alibaba-qwen-team-introduces-qwen3-5-livetranslate-flash-real-time-multimodal-interpretation-across-60-languages-at-2-8-second-latency/" }, { "name": "Command A+", "org": "Cohere", "country": "Canada", "date": "2026-05-20", "precision": "day", "category": "language", "open_weights": true, "params": "218B (25B active)", "note": "Cohere's first open-weight MoE, Apache 2.0, runs on two H100s", "source": "https://en.wikipedia.org/wiki/Cohere" }, { "name": "Qwen3.7-Max", "org": "Alibaba", "country": "China", "date": "2026-05-20", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "1M-context reasoning agent model; strongest Chinese model on AA index", "source": "https://www.marktechpost.com/2026/05/21/qwen-introduces-qwen3-7-max-a-reasoning-agent-model-with-a-1m-token-context-window/" }, { "name": "Stable Audio 3.0", "org": "Stability AI", "country": "UK", "date": "2026-05-20", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Audio model family including open-weight models", "source": "https://stability.ai/news-updates" }, { "name": "Fara1.5", "org": "Microsoft", "country": "USA", "date": "2026-05-21", "precision": "day", "category": "agent", "open_weights": true, "params": "4B / 9B / 27B", "note": "Open browser computer-use agents built on Qwen3.5", "source": "https://huggingface.co/microsoft/Fara1.5-9B" }, { "name": "Hy-MT2", "org": "Tencent", "country": "China", "date": "2026-05-21", "precision": "day", "category": "language", "open_weights": true, "params": "1.8B / 7B / 30B-A3B", "note": "Second-generation open translation models across 33 languages", "source": "https://huggingface.co/tencent/Hy-MT2-30B-A3B" }, { "name": "Lance", "org": "ByteDance", "country": "China", "date": "2026-05-21", "precision": "day", "category": "multimodal", "open_weights": true, "params": "3B active", "note": "Unified image/video understanding, generation and editing model", "source": "https://www.marktechpost.com/2026/05/21/one-model-three-modalities-bytedance-releases-lance-for-image-and-video-understanding-generation-and-editing/" }, { "name": "LongCat-Video-Avatar 1.5", "org": "Meituan", "country": "China", "date": "2026-05-21", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Audio-driven avatar video model with 8-step inference", "source": "https://news.smol.ai/issues/26-05-21-not-much" }, { "name": "StepAudio 2.5 Realtime", "org": "StepFun", "country": "China", "date": "2026-05-24", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "End-to-end realtime voice model with persona customization", "source": "https://www.marktechpost.com/2026/05/24/stepfun-releases-stepaudio-2-5-realtime-an-end-to-end-voice-model-with-roleplay-specific-rlhf-and-paralinguistic-comprehension/" }, { "name": "Keye-VL-2.0", "org": "Kuaishou", "country": "China", "date": "2026-05-25", "precision": "day", "category": "language", "open_weights": true, "params": "30B (3B active)", "note": "Long-video multimodal model with 256K context", "source": "https://github.com/Kwai-Keye/Keye" }, { "name": "Baichuan-M4", "org": "Baichuan", "country": "China", "date": "2026-05-26", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Clinical-grade medical agent system for continuous care", "source": "https://news.aibase.com/news/28350" }, { "name": "ElevenLabs Music v2", "org": "ElevenLabs", "country": "USA", "date": "2026-05-26", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Better vocals, instrumentation and multilingual support", "source": "https://elevenlabs.io/blog/introducing-music-v2" }, { "name": "Stable Audio 3", "org": "Stability AI", "country": "UK", "date": "2026-05-26", "precision": "day", "category": "audio", "open_weights": true, "params": "459M-2.7B", "note": "Fast latent-diffusion audio family; small and medium open weights", "source": "https://www.marktechpost.com/2026/05/26/stability-ai-releases-stable-audio-3-a-family-of-fast-latent-diffusion-models-for-audio-generation-and-editing/" }, { "name": "Claude Opus 4.8", "org": "Anthropic", "country": "USA", "date": "2026-05-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Same-price Opus upgrade just six weeks after 4.7", "source": "https://platform.claude.com/docs/en/release-notes/overview" }, { "name": "ElevenLabs Dubbing v2", "org": "ElevenLabs", "country": "UK", "date": "2026-05-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Upgraded AI dubbing model for video localization", "source": "https://elevenlabs.io/blog" }, { "name": "LFM2.5-8B-A1B", "org": "Liquid AI", "country": "USA", "date": "2026-05-28", "precision": "day", "category": "language", "open_weights": true, "params": "8.3B (1.5B active)", "note": "On-device MoE trained on 38T tokens", "source": "https://www.marktechpost.com/2026/05/28/liquid-ai-releases-lfm2-5-8b-a1b-an-on-device-moe-model-with-8-3b-total-and-1-5b-active-parameters/" }, { "name": "Step 3.7 Flash", "org": "StepFun", "country": "China", "date": "2026-05-28", "precision": "day", "category": "language", "open_weights": true, "params": "198B (MoE)", "note": "Open vision-language MoE for coding agents and search", "source": "https://pandaily.com/stepfun-open-source-step-3-7-flash-llm-agent-may2026" }, { "name": "Cosmos 3", "org": "NVIDIA", "country": "USA", "date": "2026-05-31", "precision": "day", "category": "robotics", "open_weights": true, "params": "64B (Super)", "note": "Omnimodal physical-AI model unifying reasoning, world and action generation", "source": "https://huggingface.co/nvidia/Cosmos3-Super" }, { "name": "Brain2Qwerty v2", "org": "Meta", "country": "USA", "date": "2026-06-01", "precision": "month", "category": "science", "open_weights": false, "params": "", "note": "Non-invasive brain-to-text decoder, ~61% word accuracy", "source": "https://news.smol.ai/issues/26-06-29-not-much" }, { "name": "Fun-Realtime-TTS", "org": "Alibaba", "country": "China", "date": "2026-06-01", "precision": "month", "category": "audio", "open_weights": false, "params": "", "note": "Debuted #1 on Artificial Analysis Speech Arena", "source": "https://news.smol.ai/issues/26-06-03-not-much" }, { "name": "Holo3.1", "org": "H Company", "country": "France", "date": "2026-06-01", "precision": "day", "category": "agent", "open_weights": true, "params": "0.8B-35B", "note": "Fast, local computer-use models", "source": "https://www.hcompany.ai/blog" }, { "name": "Krea 2", "org": "Krea", "country": "USA", "date": "2026-06-01", "precision": "month", "category": "image", "open_weights": true, "params": "", "note": "In-house open image model with Raw and Turbo variants", "source": "https://news.smol.ai/issues/26-06-23-not-much" }, { "name": "Magenta RealTime 2", "org": "Google DeepMind", "country": "UK", "date": "2026-06-01", "precision": "month", "category": "audio", "open_weights": true, "params": "", "note": "Open low-latency streaming music generator for on-device use", "source": "https://news.smol.ai/issues/26-06-03-not-much" }, { "name": "Mellum2", "org": "JetBrains", "country": "Netherlands", "date": "2026-06-01", "precision": "day", "category": "code", "open_weights": true, "params": "12B (2.5B active)", "note": "Fast open MoE for IDE, routing and RAG workloads", "source": "https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Instruct" }, { "name": "MiniMax-M3", "org": "MiniMax", "country": "China", "date": "2026-06-01", "precision": "day", "category": "multimodal", "open_weights": true, "params": "428B (23B active)", "note": "Native multimodal coding model, 1M context, MiniMax Sparse Attention", "source": "https://www.marktechpost.com/2026/06/01/minimax-releases-minimax-m3-with-msa-architecture-supporting-1m-token-context-native-multimodality-and-agentic-coding/" }, { "name": "Nemotron-TwoTower-30B-A3B", "org": "NVIDIA", "country": "USA", "date": "2026-06-01", "precision": "month", "category": "language", "open_weights": true, "params": "2x30B towers (3B active)", "note": "Block-diffusion LM with 2.4x throughput over autoregressive baseline", "source": "https://news.smol.ai/issues/26-06-25-not-much" }, { "name": "Qwen-AgentWorld-35B-A3B", "org": "Alibaba", "country": "China", "date": "2026-06-01", "precision": "month", "category": "agent", "open_weights": true, "params": "35B (3B active)", "note": "Language world model simulating MCP, terminal, Android and OS environments", "source": "https://news.smol.ai/issues/26-06-24-not-much" }, { "name": "Qwen3.7-Plus", "org": "Alibaba", "country": "China", "date": "2026-06-01", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Cost-effective model with upgraded vision-language abilities", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "MAI-Code-1-Flash", "org": "Microsoft", "country": "USA", "date": "2026-06-02", "precision": "day", "category": "code", "open_weights": false, "params": "5B", "note": "Efficient coding model for GitHub Copilot and VS Code", "source": "https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/" }, { "name": "MAI-Image-2.5", "org": "Microsoft", "country": "USA", "date": "2026-06-02", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Launched #2 for image editing on Arena", "source": "https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/" }, { "name": "MAI-Image-2.5-Flash", "org": "Microsoft", "country": "USA", "date": "2026-06-02", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Ultra-efficient variant of MAI-Image-2.5", "source": "https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/" }, { "name": "MAI-Thinking-1", "org": "Microsoft", "country": "USA", "date": "2026-06-02", "precision": "day", "category": "reasoning", "open_weights": false, "params": "MoE, 35B active", "note": "Microsoft's first in-house frontier-class reasoning model, launched at Build", "source": "https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/" }, { "name": "MAI-Transcribe-1.5", "org": "Microsoft", "country": "USA", "date": "2026-06-02", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech recognition expanded to 43 languages", "source": "https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/" }, { "name": "MAI-Voice-2", "org": "Microsoft", "country": "USA", "date": "2026-06-02", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech synthesis in 15 languages with voice adaptation", "source": "https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/" }, { "name": "MAI-Voice-2-Flash", "org": "Microsoft", "country": "USA", "date": "2026-06-02", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "2x faster, 32% cheaper MAI-Voice-2 variant", "source": "https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/" }, { "name": "Zamba2-VL", "org": "Zyphra", "country": "USA", "date": "2026-06-02", "precision": "day", "category": "language", "open_weights": true, "params": "7B", "note": "Vision-language models based on Zamba2 hybrids", "source": "https://www.zyphra.com/our-work" }, { "name": "Gemma 4 12B", "org": "Google DeepMind", "country": "UK", "date": "2026-06-03", "precision": "day", "category": "multimodal", "open_weights": true, "params": "12B", "note": "Encoder-free multimodal Gemma with native audio; runs in 16GB", "source": "https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12b/" }, { "name": "Grok Imagine Video 1.5", "org": "xAI", "country": "USA", "date": "2026-06-03", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Image-to-video model with synchronized audio; previewed Jun 3, GA Jun 16.", "source": "https://x.ai/news/grok-imagine-1-5" }, { "name": "Ideogram 4.0", "org": "Ideogram", "country": "Canada", "date": "2026-06-03", "precision": "day", "category": "image", "open_weights": true, "params": "9.3B", "note": "Ideogram goes open-weight under Apache 2.0", "source": "https://en.wikipedia.org/wiki/Ideogram_(text-to-image_model)" }, { "name": "Nemotron 3.5 ASR", "org": "NVIDIA", "country": "USA", "date": "2026-06-04", "precision": "day", "category": "audio", "open_weights": true, "params": "0.6B", "note": "Streaming multilingual ASR for 40 language-locales", "source": "https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b" }, { "name": "Nemotron 3.5 ASR Streaming", "org": "NVIDIA", "country": "USA", "date": "2026-06-04", "precision": "day", "category": "audio", "open_weights": true, "params": "0.6B", "note": "Cache-aware streaming ASR for voice agents", "source": "https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b" }, { "name": "Nemotron 3.5 Content Safety", "org": "NVIDIA", "country": "USA", "date": "2026-06-04", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Customizable multimodal safety guard for text, image and audio", "source": "https://dev.to/vjswamy/latest-ai-model-releases-june-2026-roundup-49j5" }, { "name": "ADM 3 Cloud", "org": "Apple", "country": "USA", "date": "2026-06-08", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Server image generation and editing model", "source": "https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models" }, { "name": "AFM 3 Cloud", "org": "Apple", "country": "USA", "date": "2026-06-08", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Server model on Apple Private Cloud Compute", "source": "https://en.wikipedia.org/wiki/Apple_Intelligence" }, { "name": "AFM 3 Cloud Pro", "org": "Apple", "country": "USA", "date": "2026-06-08", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Top Apple model, served on Nvidia GPUs via Google Cloud", "source": "https://en.wikipedia.org/wiki/Apple_Intelligence" }, { "name": "AFM 3 Core", "org": "Apple", "country": "USA", "date": "2026-06-08", "precision": "day", "category": "language", "open_weights": false, "params": "3B", "note": "Third-generation on-device Apple Foundation Model, WWDC 2026", "source": "https://en.wikipedia.org/wiki/Apple_Intelligence" }, { "name": "AFM 3 Core Advanced", "org": "Apple", "country": "USA", "date": "2026-06-08", "precision": "day", "category": "language", "open_weights": false, "params": "20B", "note": "Larger on-device Apple model for select devices", "source": "https://en.wikipedia.org/wiki/Apple_Intelligence" }, { "name": "Claude Fable 5", "org": "Anthropic", "country": "USA", "date": "2026-06-09", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "First public Mythos-class model; classifiers route risky topics", "source": "https://www.anthropic.com/news/claude-fable-5-mythos-5" }, { "name": "Claude Mythos 5", "org": "Anthropic", "country": "USA", "date": "2026-06-09", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Fable 5 capabilities without safety classifiers; restricted access", "source": "https://www.anthropic.com/news/claude-fable-5-mythos-5" }, { "name": "Gemini 3.5 Live Translate", "org": "Google DeepMind", "country": "UK", "date": "2026-06-09", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-to-speech translation preserving speaker intonation, 70+ languages", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-live-3-5-translate/" }, { "name": "Luma Ray3.2", "org": "Luma AI", "country": "USA", "date": "2026-06-09", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Frame-by-frame directorial control model and API", "source": "https://lumalabs.ai/news" }, { "name": "North Mini Code", "org": "Cohere", "country": "Canada", "date": "2026-06-09", "precision": "day", "category": "code", "open_weights": true, "params": "30B (3B active)", "note": "Cohere's first agentic coding model, Apache-2.0", "source": "https://cohere.com/blog/north-mini-code" }, { "name": "North Mini Code 1.0", "org": "Cohere", "country": "Canada", "date": "2026-06-09", "precision": "day", "category": "code", "open_weights": true, "params": "30B (3B active)", "note": "Apache 2.0 agentic coding MoE with 256K context", "source": "https://huggingface.co/CohereLabs/North-Mini-Code-1.0" }, { "name": "DiffusionGemma", "org": "Google DeepMind", "country": "UK", "date": "2026-06-10", "precision": "day", "category": "language", "open_weights": true, "params": "25.2B (3.8B active)", "note": "Open text-diffusion Gemma generating up to 4x faster", "source": "https://www.marktechpost.com/2026/06/10/google-ai-releases-diffusiongemma-a-26b-moe-open-model-using-text-diffusion-for-up-to-4x-faster-generation/" }, { "name": "Kimi K2.7 Code", "org": "Moonshot AI", "country": "China", "date": "2026-06-12", "precision": "day", "category": "code", "open_weights": true, "params": "1T (32B active)", "note": "Coding specialist using ~30% fewer reasoning tokens than K2.6", "source": "https://www.marktechpost.com/2026/06/12/moonshot-ai-releases-kimi-k2-7-code-a-coding-model-reporting-21-8-on-kimi-code-bench-v2-over-k2-6/" }, { "name": "openPangu 2.0", "org": "Huawei", "country": "China", "date": "2026-06-12", "precision": "day", "category": "language", "open_weights": true, "params": "505B", "note": "Open Pangu 2.0 family with 512K context; Pro weights out July 31", "source": "https://tech.ifeng.com/c/8ttgqM0UZSp" }, { "name": "Pangu 6.0", "org": "Huawei", "country": "China", "date": "2026-06-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Embedded in the HarmonyOS 7 kernel", "source": "https://tech.ifeng.com/c/8ttndK7Q0Az" }, { "name": "Physis", "org": "BAAI", "country": "China", "date": "2026-06-12", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "General world foundation model unveiled at BAAI Conference 2026", "source": "https://www.wedoany.com/shortnews/251491.html" }, { "name": "Spark X2-VL", "org": "iFlytek", "country": "China", "date": "2026-06-12", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Multimodal model trained on domestic compute", "source": "https://news.mydrivers.com/1/1129/1129359.htm" }, { "name": "ZONOS2", "org": "Zyphra", "country": "USA", "date": "2026-06-12", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Second-generation Zonos text-to-speech", "source": "https://www.zyphra.com/our-work" }, { "name": "GLM-5.2", "org": "Zhipu AI", "country": "China", "date": "2026-06-13", "precision": "day", "category": "language", "open_weights": true, "params": "744B (40B active)", "note": "1M context; top open model on frontend-coding arenas", "source": "https://www.marktechpost.com/2026/06/14/z-ai-launches-glm-5-2-with-a-usable-1m-token-context-two-thinking-effort-levels-and-no-benchmarks-at-launch/" }, { "name": "BoltzMol-1", "org": "Boltz", "country": "USA", "date": "2026-06-16", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Small-molecule screening and hit-discovery model via Boltz API", "source": "https://boltz.com/news" }, { "name": "BoltzProt-1", "org": "Boltz", "country": "USA", "date": "2026-06-16", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Protein design model launched with the Boltz API", "source": "https://boltz.com/news" }, { "name": "HappyHorse-1.1", "org": "Alibaba", "country": "China", "date": "2026-06-16", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Update to Alibaba's HappyHorse text/image/reference-to-video models", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Qwen-RobotSuite", "org": "Alibaba", "country": "China", "date": "2026-06-16", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "VLA manipulation, video world-model and navigation models", "source": "https://www.marktechpost.com/2026/06/16/meet-qwen-robotsuite-three-embodied-ai-models-for-vla-manipulation-video-world-modeling-and-navigation/" }, { "name": "MolmoMotion", "org": "Allen Institute for AI", "country": "USA", "date": "2026-06-17", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Language-guided 3D motion forecasting on Molmo 2 backbone", "source": "https://allenai.org/blog/molmo-motion" }, { "name": "LFM2.5 Retrievers", "org": "Liquid AI", "country": "USA", "date": "2026-06-18", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Bidirectional LFMs for fast multilingual search", "source": "https://www.liquid.ai/blog/lfm2-5-retrievers" }, { "name": "Unlimited-OCR", "org": "Baidu", "country": "China", "date": "2026-06-22", "precision": "day", "category": "vision", "open_weights": true, "params": "3B", "note": "MIT-licensed OCR keeping KV cache flat for long documents", "source": "https://huggingface.co/baidu/Unlimited-OCR" }, { "name": "Mistral OCR 4", "org": "Mistral AI", "country": "France", "date": "2026-06-23", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Document-intelligence OCR with bounding boxes, 170 languages", "source": "https://mistral.ai/news" }, { "name": "Seed 2.1 Pro", "org": "ByteDance", "country": "China", "date": "2026-06-23", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Flagship Seed2.1 model focused on AI productivity", "source": "https://seed.bytedance.com/blog/seed2-1-officially-released-advancing-ai-productivity" }, { "name": "Seed 2.1 Turbo", "org": "ByteDance", "country": "China", "date": "2026-06-23", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Faster, cheaper Seed2.1 variant", "source": "https://seed.bytedance.com/blog/seed2-1-officially-released-advancing-ai-productivity" }, { "name": "Gradium STT-Translate / S2S-Translate", "org": "Gradium", "country": "France", "date": "2026-06-24", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Single-pass streaming speech translation models", "source": "https://www.marktechpost.com/2026/06/24/gradium-launches-stt-translate-and-s2s-translate-real-time-speech-translation-models-beating-gpt-realtime-translate-on-accuracy-and-latency/" }, { "name": "LFM2.5-230M", "org": "Liquid AI", "country": "USA", "date": "2026-06-25", "precision": "day", "category": "language", "open_weights": true, "params": "230M", "note": "Smallest LFM2.5 model built to run anywhere", "source": "https://www.liquid.ai/blog/lfm2-5-230m" }, { "name": "GPT-5.6 Luna", "org": "OpenAI", "country": "USA", "date": "2026-06-26", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Fastest, cheapest GPT-5.6 tier", "source": "https://news.smol.ai/issues/26-06-26-gpt-56-preview" }, { "name": "GPT-5.6 Sol", "org": "OpenAI", "country": "USA", "date": "2026-06-26", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Flagship previewed under new US government pre-release review", "source": "https://en.wikipedia.org/wiki/GPT-5.6" }, { "name": "GPT-5.6 Terra", "org": "OpenAI", "country": "USA", "date": "2026-06-26", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Balanced tier at roughly half GPT-5.5's price", "source": "https://news.smol.ai/issues/26-06-26-gpt-56-preview" }, { "name": "LongCat-2.0", "org": "Meituan", "country": "China", "date": "2026-06-29", "precision": "day", "category": "code", "open_weights": true, "params": "1.6T", "note": "Near-frontier agentic coding model trained entirely on Chinese chips", "source": "https://venturebeat.com/technology/meituan-open-sources-longcat-2-0-the-1-6t-near-frontier-agentic-coding-model-thats-been-leading-openrouter-trained-entirely-on-chinese-chips" }, { "name": "Claude Sonnet 5", "org": "Anthropic", "country": "USA", "date": "2026-06-30", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Most agentic Sonnet; near Opus 4.8 at far lower cost", "source": "https://www.anthropic.com/news/claude-sonnet-5" }, { "name": "Leanstral 1.5", "org": "Mistral AI", "country": "France", "date": "2026-06-30", "precision": "day", "category": "science", "open_weights": true, "params": "119B (6B active)", "note": "Upgraded open theorem-proving model for Lean 4", "source": "https://docs.mistral.ai/getting-started/changelog/" }, { "name": "Nano Banana 2 Lite", "org": "Google DeepMind", "country": "UK", "date": "2026-06-30", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Fastest, cheapest Gemini image model, about 4-second generations", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni-flash-nano-banana-2-lite/" }, { "name": "TabFM", "org": "Google", "country": "USA", "date": "2026-06-30", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Zero-shot foundation model for tabular data prediction", "source": "https://research.google/blog/introducing-tabfm-a-zero-shot-foundation-model-for-tabular-data/" }, { "name": "Agents-A1", "org": "Shanghai AI Lab", "country": "China", "date": "2026-07-01", "precision": "day", "category": "agent", "open_weights": true, "params": "35B", "note": "Open 35B agent model claimed to rival trillion-parameter models", "source": "https://eu.36kr.com/en/p/3877948838244353" }, { "name": "Apertus 1.5", "org": "Swiss AI Initiative", "country": "Switzerland", "date": "2026-07-01", "precision": "month", "category": "language", "open_weights": true, "params": "8B / 70B", "note": "Updated fully open Swiss models", "source": "https://en.wikipedia.org/wiki/Apertus_(LLM)" }, { "name": "GigaChat 3.5 Ultra", "org": "Sber", "country": "Russia", "date": "2026-07-01", "precision": "month", "category": "language", "open_weights": true, "params": "432B (28B active)", "note": "Smaller yet stronger open flagship with reasoning variant", "source": "https://huggingface.co/ai-sage/GigaChat3.5-432B-A28B" }, { "name": "Kandinsky WM 1.0", "org": "Sber", "country": "Russia", "date": "2026-07-01", "precision": "month", "category": "video", "open_weights": true, "params": "", "note": "Sber's image-to-video world model", "source": "https://huggingface.co/api/models?author=kandinskylab&sort=createdAt&direction=-1&limit=60" }, { "name": "Vidu S1", "org": "Shengshu", "country": "China", "date": "2026-07-03", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Real-time interactive video generation model", "source": "https://www.leiphone.com/category/industrynews/6GlFzI5hMwcfRoGZ.html" }, { "name": "GPT-Realtime-2.1", "org": "OpenAI", "country": "USA", "date": "2026-07-06", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Realtime reasoning voice model with better alphanumeric recognition", "source": "https://datanorth.ai/news/openai-releases-gpt-realtime-2-1-voice-models" }, { "name": "GPT-Realtime-2.1 mini", "org": "OpenAI", "country": "USA", "date": "2026-07-06", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Distilled, cheaper realtime reasoning voice model", "source": "https://datanorth.ai/news/openai-releases-gpt-realtime-2-1-voice-models" }, { "name": "Hy3", "org": "Tencent", "country": "China", "date": "2026-07-06", "precision": "day", "category": "language", "open_weights": true, "params": "295B", "note": "Tencent's open-source flagship MoE LLM, Apache 2.0.", "source": "https://the-decoder.com/tencent-releases-hy3-open-source-model-that-allegedly-matches-models-up-to-five-times-its-active-size/" }, { "name": "MIRA World Model", "org": "Kyutai", "country": "France", "date": "2026-07-06", "precision": "day", "category": "video", "open_weights": false, "params": "5B", "note": "Real-time Rocket League world model with General Intuition and Epic", "source": "https://mira-wm.com/blog-post/" }, { "name": "Cohere Transcribe Arabic", "org": "Cohere", "country": "Canada", "date": "2026-07-07", "precision": "day", "category": "audio", "open_weights": true, "params": "2B", "note": "Open Arabic speech recognition handling dialects and code-switching.", "source": "https://the-decoder.com/cohere-transcribe-arabic-is-an-open-source-model-built-for-arabics-toughest-transcription-problems/" }, { "name": "Muse Image", "org": "Meta", "country": "USA", "date": "2026-07-07", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "First media-generation model from Meta Superintelligence Labs; agentic image generation.", "source": "https://ai.meta.com/blog/introducing-muse-image-muse-video-msl/" }, { "name": "Muse Video", "org": "Meta", "country": "USA", "date": "2026-07-07", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Early preview of Meta's video generator with native audio.", "source": "https://ai.meta.com/blog/introducing-muse-image-muse-video-msl/" }, { "name": "GPT-Live-1", "org": "OpenAI", "country": "USA", "date": "2026-07-08", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Full-duplex ChatGPT voice model that listens while speaking; mini variant for free users.", "source": "https://cryptobriefing.com/openai-gpt-live-1-full-duplex-voice-model/" }, { "name": "Grok 4.5", "org": "xAI", "country": "USA", "date": "2026-07-08", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Coding and agent model trained alongside Cursor at low prices.", "source": "https://x.ai/news/grok-4-5" }, { "name": "Robostral Navigate", "org": "Mistral AI", "country": "France", "date": "2026-07-08", "precision": "day", "category": "robotics", "open_weights": false, "params": "8B", "note": "Mistral's first robotics model; steers robots with one camera.", "source": "https://the-decoder.com/mistral-enters-robotics-with-robostral-navigate-an-8b-model-that-steers-robots-using-just-one-camera/" }, { "name": "Seedream 5.0 Pro", "org": "ByteDance", "country": "China", "date": "2026-07-08", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Design-aware image model with advanced reasoning", "source": "https://seed.bytedance.com/blog/beyond-generation-it-understands-design-introducing-seedream-5-0-pro" }, { "name": "SWE-1.7", "org": "Cognition", "country": "USA", "date": "2026-07-08", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Cognition's in-house RL-trained coding model for Devin and Windsurf.", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Muse Spark 1.1", "org": "Meta", "country": "USA", "date": "2026-07-09", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Meta's first paid developer API model, aimed at agentic coding.", "source": "https://techcrunch.com/2026/07/09/meta-enters-the-crowded-ai-coding-battle-with-muse-spark-1-1/" }, { "name": "MuScriptor", "org": "Kyutai", "country": "France", "date": "2026-07-10", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Automatic multi-instrument music transcription model", "source": "https://kyutai.org/blog" }, { "name": "Step Edge", "org": "StepFun", "country": "China", "date": "2026-07-12", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "High-efficiency multimodal models for edge devices", "source": "https://static.stepfun.com/blog/step-edge/" }, { "name": "SensorFM", "org": "Google", "country": "USA", "date": "2026-07-13", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Wearable-sensor foundation model trained on a trillion minutes of data.", "source": "https://the-decoder.com/sensorfm/" }, { "name": "Soofi S", "org": "KI Bundesverband consortium", "country": "Germany", "date": "2026-07-13", "precision": "day", "category": "language", "open_weights": true, "params": "31.6B (3.2B active)", "note": "Fully open German-English MoE meeting the OSI open-source AI definition.", "source": "https://the-decoder.com/german-ai-consortium-releases-soofi-s-an-open-30b-model-that-tops-benchmarks-in-both-english-and-german/" }, { "name": "Qwen-Audio-3.0", "org": "Alibaba", "country": "China", "date": "2026-07-14", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "New TTS family, followed by ASR and realtime duplex models", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "SenseNova-Vision", "org": "SenseTime", "country": "China", "date": "2026-07-14", "precision": "day", "category": "vision", "open_weights": true, "params": "", "note": "Open unified vision model covering four core vision tasks", "source": "https://technode.com/2026/07/14/sensetime-open-sources-sensenova-vision-unified-vision-model/" }, { "name": "Bonsai 27B", "org": "PrismML", "country": "USA", "date": "2026-07-15", "precision": "day", "category": "reasoning", "open_weights": true, "params": "27B", "note": "1-2 bit compressed open reasoning model that runs on an iPhone.", "source": "https://the-decoder.com/bonsai-27b-is-a-full-open-reasoning-model-that-fits-on-an-iphone/" }, { "name": "Inkling", "org": "Thinking Machines Lab", "country": "USA", "date": "2026-07-15", "precision": "day", "category": "language", "open_weights": true, "params": "975B", "note": "Mira Murati's lab ships its first model, as open weights.", "source": "https://the-decoder.com/ex-openai-cto-muratis-thinking-machines-drops-inkling-a-975b-parameter-model-that-leads-us-labs-but-trails-china/" }, { "name": "Cosmos 3 Edge", "org": "NVIDIA", "country": "USA", "date": "2026-07-16", "precision": "day", "category": "robotics", "open_weights": true, "params": "4B", "note": "Compact on-device world model for robot perception and action.", "source": "https://adtmag.com/articles/2026/07/16/nvidia-expands-cosmos-physical-ai-platform-with-edge-model.aspx" }, { "name": "Kimi K3", "org": "Moonshot AI", "country": "China", "date": "2026-07-16", "precision": "day", "category": "language", "open_weights": true, "params": "2.8T", "note": "Largest open-weight model to date; weights published July 26.", "source": "https://en.wikipedia.org/wiki/Kimi_(chatbot)" }, { "name": "Music-3.0", "org": "MiniMax", "country": "China", "date": "2026-07-16", "precision": "day", "category": "audio", "open_weights": true, "params": "2B", "note": "New-generation music model with weights on Hugging Face", "source": "https://platform.minimax.io/docs/release-notes/models" }, { "name": "ZUNA1.1", "org": "Zyphra", "country": "USA", "date": "2026-07-16", "precision": "day", "category": "science", "open_weights": true, "params": "400M", "note": "Updated brain-computer-interface foundation model", "source": "https://www.zyphra.com/our-work" }, { "name": "Qwen3.8-Max", "org": "Alibaba", "country": "China", "date": "2026-07-19", "precision": "day", "category": "language", "open_weights": true, "params": "2.4T (95B active)", "note": "Alibaba's largest model; API August 3, open weights August 12.", "source": "https://the-decoder.com/alibabas-qwen-takes-on-kimi-k3-with-open-weight-qwen-3-8-says-model-is-second-only-to-fable-5/" }, { "name": "SenseNova U1 Pro", "org": "SenseTime", "country": "China", "date": "2026-07-19", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Agentic multimodal generation model with up to 8K output, launched at WAIC", "source": "https://www.moomoo.com/news/post/73183301/hk-stock-market-movement-sensetime-w-00020-hk-rises-over" }, { "name": "Seed Audio 1.0", "org": "ByteDance", "country": "China", "date": "2026-07-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Audio creation model", "source": "https://seed.bytedance.com/en/seedaudio1_0" }, { "name": "Fugu-Cyber", "org": "Sakana AI", "country": "Japan", "date": "2026-07-21", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Orchestration model with state-of-the-art cybersecurity benchmark results", "source": "https://sakana.ai/blog/" }, { "name": "Gemini 3.5 Flash Cyber", "org": "Google DeepMind", "country": "UK", "date": "2026-07-21", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Vulnerability-hunting Gemini fine-tune in limited pilot via CodeMender.", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/" }, { "name": "Gemini 3.5 Flash-Lite", "org": "Google DeepMind", "country": "UK", "date": "2026-07-21", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Fastest, cheapest 3.5-class Gemini at 350 tokens per second.", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/" }, { "name": "Gemini 3.6 Flash", "org": "Google DeepMind", "country": "UK", "date": "2026-07-21", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Workhorse Flash; first of three Gemini Flash releases in six weeks.", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/" }, { "name": "Laguna S 2.1", "org": "Poolside", "country": "USA", "date": "2026-07-21", "precision": "day", "category": "code", "open_weights": true, "params": "118B", "note": "Open-weight agentic coding model small enough for one desktop.", "source": "https://the-decoder.com/poolsides-laguna-s-2-1-is-a-small-open-weight-coding-model-that-punches-well-above-its-size/" }, { "name": "Motif 3", "org": "Motif Technologies", "country": "South Korea", "date": "2026-07-21", "precision": "day", "category": "language", "open_weights": true, "params": "314B", "note": "Korean sovereign-AI contender released with MIT-licensed weights.", "source": "https://en.wikipedia.org/wiki/List_of_large_language_models" }, { "name": "Qwen-Audio-3.0-TTS-Plus", "org": "Alibaba", "country": "China", "date": "2026-07-21", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Topped Artificial Analysis Speech Arena text-to-speech leaderboard.", "source": "https://the-decoder.com/alibabas-qwen-audio-3-0-tts-plus-tops-the-competition-in-the-text-to-speech-rankings/" }, { "name": "Qwen-Image-3.0", "org": "Alibaba", "country": "China", "date": "2026-07-21", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Renders full infographic grids and readable tiny text in one pass.", "source": "https://the-decoder.com/alibabas-qwen-image-3-0-renders-full-infographic-grids-and-readable-ten-pixel-text-in-a-single-pass/" }, { "name": "Qwen3.7-Flash", "org": "Alibaba", "country": "China", "date": "2026-07-21", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Fast tier with enhanced multimodal understanding and agent execution", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Xiaomi-Robotics-1", "org": "Xiaomi", "country": "China", "date": "2026-07-21", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Manipulation model trained on 100,000+ hours of handheld-gripper data.", "source": "https://the-decoder.com/xiaomi-robotics-1-shows-that-more-data-beats-bigger-models-when-training-robots-to-move/" }, { "name": "Antares-1B", "org": "Cisco", "country": "USA", "date": "2026-07-22", "precision": "day", "category": "code", "open_weights": true, "params": "1B", "note": "Small open vulnerability-detection models (1B and 350M) for local use.", "source": "https://the-decoder.com/cisco-bets-its-small-open-cybersecurity-models-can-outperform-gpt-5-5-at-vulnerability-detection-for-a-fraction-of-the-cost/" }, { "name": "Genesis-Science-1", "org": "Arcee AI", "country": "USA", "date": "2026-07-22", "precision": "day", "category": "science", "open_weights": false, "params": "~1T", "note": "Trillion-class Trinity-based science model; open weights promised later", "source": "https://www.arcee.ai/blog/genesis-science-1" }, { "name": "Solar Open 2", "org": "Upstage", "country": "South Korea", "date": "2026-07-22", "precision": "day", "category": "language", "open_weights": true, "params": "250B (15B active)", "note": "Korean sovereign hybrid-attention MoE with 1M context", "source": "https://www.upstage.ai/blog" }, { "name": "FLUX 3", "org": "Black Forest Labs", "country": "Germany", "date": "2026-07-23", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Unified image, video, audio and action model; 20-second videos with audio.", "source": "https://www.techtimes.com/articles/321552/20260725/flux-3-launches-black-forest-labs-enters-video-audio-physical-ai-one-model.htm" }, { "name": "FLUX-mimic", "org": "Black Forest Labs", "country": "Germany", "date": "2026-07-23", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Video-action model with mimic robotics for dexterous manipulation", "source": "https://bfl.ai/blog/flux-3" }, { "name": "Ling-3.0-flash", "org": "Ant Group", "country": "China", "date": "2026-07-23", "precision": "day", "category": "language", "open_weights": true, "params": "124B", "note": "Hybrid-reasoning open MoE from Ant's Bailing team.", "source": "https://cryptobriefing.com/ant-group-ling-3-flash-124b-open-weights/" }, { "name": "MAI-Image-2.5-Pro", "org": "Microsoft", "country": "USA", "date": "2026-07-23", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Maximum-fidelity image model for hero imagery and text", "source": "https://microsoft.ai/news/introducing-mai-image-2-5-pro-and-mai-voice-2-flash/" }, { "name": "Claude Opus 5", "org": "Anthropic", "country": "USA", "date": "2026-07-24", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Near-Fable 5 performance at half the price; effort toggle.", "source": "https://www.anthropic.com/news/claude-opus-5" }, { "name": "Fugu Ultra v1.1", "org": "Sakana AI", "country": "Japan", "date": "2026-07-24", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Multi-model orchestration system claimed to beat Claude Fable 5.", "source": "https://the-decoder.com/sakana-claims-its-ai-model-router-fugu-ultra-v1-1-now-beats-fable-5-without-even-including-it-in-the-pool/" }, { "name": "Instella-MoE", "org": "AMD", "country": "USA", "date": "2026-07-24", "precision": "day", "category": "language", "open_weights": true, "params": "16B (2.8B active)", "note": "Fully open MoE trained on MI300X and MI325X", "source": "https://rocm.blogs.amd.com/artificial-intelligence/instella-moe/README.html" }, { "name": "Midjourney V8.2", "org": "Midjourney", "country": "USA", "date": "2026-07-24", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Latest V8-series update", "source": "https://en.wikipedia.org/wiki/Midjourney" }, { "name": "KAT-Coder-V2.5", "org": "Kuaishou", "country": "China", "date": "2026-07-26", "precision": "day", "category": "code", "open_weights": true, "params": "", "note": "Agentic coding model trained on 100k+ repo environments", "source": "https://hackernoon.com/kat-coder-v25-dev-an-open-agentic-coding-model" }, { "name": "MAI-Cyber-1-Flash", "org": "Microsoft", "country": "USA", "date": "2026-07-27", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Microsoft's first cybersecurity-specialized model.", "source": "https://techcrunch.com/2026/07/27/microsoft-launches-its-first-cyber-model-and-a-new-agentic-cybersecurity-system/" }, { "name": "GPT Live Transcribe", "org": "OpenAI", "country": "USA", "date": "2026-07-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Low-latency streaming transcription model", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "GPT Transcribe", "org": "OpenAI", "country": "USA", "date": "2026-07-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "File transcription with context and keyword hints; Whisper successor", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "LFM2.5-Encoders", "org": "Liquid AI", "country": "USA", "date": "2026-07-28", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Long-context encoders fast even on CPU", "source": "https://www.liquid.ai/blog/lfm2-5-encoders" }, { "name": "A.X K2", "org": "SK Telecom", "country": "South Korea", "date": "2026-07-29", "precision": "day", "category": "language", "open_weights": true, "params": "688B (33B active)", "note": "Korean sovereign foundation model, Apache 2.0 on Hugging Face.", "source": "https://biz.chosun.com/en/en-it/2026/07/29/X6RUYHAE25BWREQ2XMRA7IBJLA/" }, { "name": "Grok Voice Think Fast 2.0", "org": "xAI", "country": "USA", "date": "2026-07-29", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Next-gen speech-to-speech model with better reasoning and transcription accuracy.", "source": "https://x.ai/news/grok-voice-think-fast-2" }, { "name": "Lyria 3.5", "org": "Google DeepMind", "country": "UK", "date": "2026-07-29", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music model with better vocals and lyrics; powers Flow Music.", "source": "https://blog.google/innovation-and-ai/models-and-research/google-labs/lyria-3-5/" }, { "name": "Gemini Robotics 2", "org": "Google DeepMind", "country": "UK", "date": "2026-07-30", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "VLA model with whole-body humanoid control, feet to fingertips.", "source": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/" }, { "name": "Gemini Robotics ER 2", "org": "Google DeepMind", "country": "UK", "date": "2026-07-30", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Embodied-reasoning robot brain with multi-robot collaboration.", "source": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/" }, { "name": "Gemini Robotics On-Device 2", "org": "Google DeepMind", "country": "UK", "date": "2026-07-30", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Local VLA that adapts to new robot bodies in hours.", "source": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/" }, { "name": "DeepSeek-V4-Flash-0731", "org": "DeepSeek", "country": "China", "date": "2026-07-31", "precision": "day", "category": "language", "open_weights": true, "params": "284B (13B active)", "note": "MIT-licensed Flash upgrade matching GPT-5.6 Luna at lower cost.", "source": "https://the-decoder.com/new-deepseek-flash-model-matches-openais-gpt-5-6-luna-at-roughly-60-percent-lower-cost/" }, { "name": "Inkling Small", "org": "Thinking Machines Lab", "country": "USA", "date": "2026-07-31", "precision": "day", "category": "language", "open_weights": true, "params": "276B (12B active)", "note": "Near-Inkling performance at under a third of the parameters.", "source": "https://the-decoder.com/thinking-machines-bets-on-efficiency-over-size-with-its-second-model-inkling-small/" }, { "name": "K-EXAONE 2.0", "org": "LG AI Research", "country": "South Korea", "date": "2026-07-31", "precision": "day", "category": "language", "open_weights": true, "params": "750B (37B active)", "note": "Korea's largest AI model, released on Hugging Face.", "source": "https://www.koreatimes.co.kr/business/tech-science/20260731/lg-unveils-750-bil-parameter-frontier-ai-model-k-exaone-20" }, { "name": "MiniMax H3", "org": "MiniMax", "country": "China", "date": "2026-07-31", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "First open model to top an AI video ranking.", "source": "https://www.reuters.com/world/china/chinas-minimax-releases-h3-video-model-2026-07-31/" }, { "name": "openPangu-2.0-Pro", "org": "Huawei", "country": "China", "date": "2026-07-31", "precision": "day", "category": "language", "open_weights": true, "params": "505B", "note": "Pretrained on Huawei's own chips; weights and report released.", "source": "https://www.techtimes.com/articles/322606/20260801/huawei-pangu-pro-trains-505-billion-parameters-without-nvidia-supply-chain-tells-different-story.htm" }, { "name": "Qwen3.7-Text-Embedding", "org": "Alibaba", "country": "China", "date": "2026-07-31", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Multilingual text embedding model based on Qwen3.7", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Seedance 2.5", "org": "ByteDance", "country": "China", "date": "2026-07-31", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "One-take creation with flexible multi-reference input", "source": "https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5" }, { "name": "Runway Ruby", "org": "Runway", "country": "USA", "date": "2026-08-01", "precision": "month", "category": "video", "open_weights": false, "params": "", "note": "SDR-to-HDR video conversion and grading model", "source": "https://runway.com/product/ruby" }, { "name": "GAIA-4", "org": "Wayve", "country": "UK", "date": "2026-08-03", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Multimodal world model for closed-loop driving simulation", "source": "https://wayve.ai/thinking/gaia-4/" }, { "name": "Sakana Namazu", "org": "Sakana AI", "country": "Japan", "date": "2026-08-03", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Japanese-specialized LLM API adapted from Moonshot's Kimi K2.6", "source": "https://sakana.ai/namazu-api/" }, { "name": "SenseNova U1.5-Lite-Preview", "org": "SenseTime", "country": "China", "date": "2026-08-03", "precision": "day", "category": "image", "open_weights": true, "params": "8B", "note": "Open lightweight unified model with native 4K output", "source": "https://www.leiphone.com/category/industrynews/qqTUnzcUVPuJaEeA.html" }, { "name": "SwanTale", "org": "ByteDance", "country": "China", "date": "2026-08-03", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Unified generator for speech, sound effects and music (research paper).", "source": "https://www.msn.com/en-us/news/technology/bytedance-unified-audio-ai-collapses-voice-sound-and-music-into-one-model-swantale/ar-AA29p5Y1" }, { "name": "Alpamayo 2 Super", "org": "NVIDIA", "country": "USA", "date": "2026-08-04", "precision": "day", "category": "robotics", "open_weights": true, "params": "", "note": "Larger Alpamayo driving VLA with six-camera input", "source": "https://huggingface.co/nvidia/Alpamayo2-Super" }, { "name": "LFM2.5-2.6B", "org": "Liquid AI", "country": "USA", "date": "2026-08-04", "precision": "day", "category": "language", "open_weights": true, "params": "2.6B", "note": "On-device model aimed at deploying agents everywhere", "source": "https://www.liquid.ai/blog/lfm2-5-2-6b" }, { "name": "Shieldstral", "org": "Mistral AI", "country": "France", "date": "2026-08-04", "precision": "day", "category": "language", "open_weights": true, "params": "3B", "note": "Open multimodal safety classifier checking content against plain-language policies.", "source": "https://siliconangle.com/2026/08/05/mistral-introduces-shieldstral-provide-lightweight-policy-aware-moderation-ai-models/" }, { "name": "Muse Spark 1.2", "org": "Meta", "country": "USA", "date": "2026-08-05", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "1M-context update that powers Meta's Muse Code agent.", "source": "https://pulse2.com/meta-launches-muse-spark-1-2-and-muse-code-coding-agent-with-1-million-token-context-window/" }, { "name": "SeedRealtime", "org": "ByteDance", "country": "China", "date": "2026-08-05", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Native full-duplex audio-video model shipped in the Doubao app.", "source": "https://technode.com/2026/08/05/bytedance-launches-seedrealtime-full-duplex-audio-video-model/" }, { "name": "WeatherNext 2-mini", "org": "Google DeepMind", "country": "UK", "date": "2026-08-06", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Compact open weather model that runs on one TPU in Colab.", "source": "https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/" }, { "name": "WeatherNext Cyclones", "org": "Google DeepMind", "country": "UK", "date": "2026-08-06", "precision": "day", "category": "science", "open_weights": true, "params": "", "note": "Adds about a day of cyclone forecast skill; published in Nature.", "source": "https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/" }, { "name": "Bulbul V4", "org": "Sarvam AI", "country": "India", "date": "2026-08-07", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Text-to-speech research preview announced at Sarvam Epoch", "source": "https://www.sarvam.ai/epoch/summary" }, { "name": "Grok Imagine Image 2.0", "org": "xAI", "country": "USA", "date": "2026-08-07", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Precise image generation and editing; strong typography, five-image references.", "source": "https://x.ai/news/grok-imagine-image-2" }, { "name": "Saaras V4", "org": "Sarvam AI", "country": "India", "date": "2026-08-07", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Five output modes from one speech model, 22 Indian languages", "source": "https://www.sarvam.ai/epoch/summary" }, { "name": "Sarvam Vision 2.0", "org": "Sarvam AI", "country": "India", "date": "2026-08-07", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "OCR and Indic handwriting model announced at Sarvam Epoch", "source": "https://www.sarvam.ai/epoch/summary" }, { "name": "GPT-5.6-Cyber", "org": "OpenAI", "country": "USA", "date": "2026-08-10", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Permissive cyber variant of GPT-5.6 Sol for vetted defenders.", "source": "https://the-decoder.com/openai-launches-gpt-5-6-cyber-to-help-defenders-find-vulnerabilities-before-attackers-do/" }, { "name": "MAI-Image-2.6", "org": "Microsoft", "country": "USA", "date": "2026-08-10", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Launched No. 2 on Arena text-to-image; Flash variant", "source": "https://microsoft.ai/news/mai-image-2-6-launches-at-no-2-on-arena-ahead-of-google-meta-and-xai/" }, { "name": "Muse Glimmer", "org": "Meta", "country": "USA", "date": "2026-08-10", "precision": "day", "category": "language", "open_weights": true, "params": "30B", "note": "Meta returns to open weights; distilled from Muse Spark 1.2.", "source": "https://www.indiatvnews.com/technology/news/meta-s-new-muse-glimmer-ai-model-brings-powerful-ai-agents-to-consumer-pcs-2026-08-10-1050946" }, { "name": "LTX-2.5", "org": "Lightricks", "country": "Israel", "date": "2026-08-11", "precision": "day", "category": "video", "open_weights": true, "params": "", "note": "Open world-model-oriented video release for robotics", "source": "https://en.wikipedia.org/wiki/LTX_(world_model)" }, { "name": "MAI-Code-1.1-Flash", "org": "Microsoft", "country": "USA", "date": "2026-08-11", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Better coding model at a quarter of the cost", "source": "https://microsoft.ai/news/mai-code-1-1-flash-br-better-faster-at-a-quarter-of-the-cost/" }, { "name": "Nemotron 3.5 Lightning", "org": "NVIDIA", "country": "USA", "date": "2026-08-11", "precision": "day", "category": "language", "open_weights": true, "params": "30B (MoE)", "note": "Efficiency-first open model for long-running multi-agent workloads.", "source": "https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/" }, { "name": "Solar Pro 4", "org": "Upstage", "country": "South Korea", "date": "2026-08-11", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Agentic model for multi-step work and tool calling", "source": "https://www.upstage.ai/blog" }, { "name": "Grok 4.6", "org": "xAI", "country": "USA", "date": "2026-08-12", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Matched GPT-5.6 Sol on benchmarks while undercutting its price.", "source": "https://the-decoder.com/spacexais-grok-4-6-matches-openais-best-model-and-undercuts-it-on-price/" }, { "name": "LFM2.5-VL-3B", "org": "Liquid AI", "country": "USA", "date": "2026-08-12", "precision": "day", "category": "language", "open_weights": true, "params": "3B", "note": "Faster, better edge vision-language model", "source": "https://www.liquid.ai/blog/lfm2-5-vl-3b" }, { "name": "SL2T", "org": "Google DeepMind", "country": "UK", "date": "2026-08-12", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Multilingual sign-language-to-text model shipping on Pixel 11.", "source": "https://deepmind.google/blog/putting-sign-language-ai-into-users-hands/" }, { "name": "DeepSeek-V4-Pro-0813", "org": "DeepSeek", "country": "China", "date": "2026-08-13", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Production V4-Pro build with major agent gains; weights not yet released.", "source": "https://the-decoder.com/deepseek-launches-an-improved-v4-pro-model-raises-api-prices-and-makes-its-agent-software-open-source/" }, { "name": "Gemini 3.7 Flash", "org": "Google DeepMind", "country": "UK", "date": "2026-08-13", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Arrived 23 days after 3.6 Flash, at half its price.", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/" }, { "name": "MiniMax Music 3.0", "org": "MiniMax", "country": "China", "date": "2026-08-13", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open-weights production-ready music model", "source": "https://www.minimax.io/blog/minimax-music-3-0-next-generation-open-weights-production-ready-versatile-music-model" }, { "name": "Wan3.0", "org": "Alibaba", "country": "China", "date": "2026-08-13", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Up to 30-second videos in one pass from any input", "source": "https://www.alibabacloud.com/blog/wan3-0-30-second-ai-video-generation-from-any-input_603452" }, { "name": "ElevenLabs Music v2.5", "org": "ElevenLabs", "country": "USA", "date": "2026-08-14", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Most advanced ElevenLabs music model", "source": "https://elevenlabs.io/docs/changelog" }, { "name": "GLM-5.3", "org": "Zhipu AI", "country": "China", "date": "2026-08-14", "precision": "day", "category": "language", "open_weights": true, "params": "744B", "note": "Top open coding model; weights briefly held over cyber capability.", "source": "https://the-decoder.com/zhipu-ai-releases-glm-5-3-claims-its-the-strongest-open-weights-coding-model/" }, { "name": "MiniMax-Music3", "org": "MiniMax", "country": "China", "date": "2026-08-14", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Text-to-song model with vocals, tracks up to five minutes.", "source": "https://gigazine.net/gsc_news/en/20260814-minimax-music3/" }, { "name": "Qwen3.8-27B", "org": "Alibaba", "country": "China", "date": "2026-08-14", "precision": "day", "category": "language", "open_weights": true, "params": "27B", "note": "Apache-2.0 dense model bringing near-frontier coding to home PCs.", "source": "https://the-decoder.com/alibabas-qwen-team-releases-qwen-3-8-models-with-open-weights-under-the-apache-2-0-license/" }, { "name": "Pika Music", "org": "Pika", "country": "USA", "date": "2026-08-18", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Music generation model in the Pika Audio family", "source": "https://pika.art/blog" }, { "name": "Pika SFX", "org": "Pika", "country": "USA", "date": "2026-08-18", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Real-time sound-effect generation from text", "source": "https://pika.art/blog" }, { "name": "Pika Soundtrack", "org": "Pika", "country": "USA", "date": "2026-08-18", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Motion-aware synchronized soundscapes for video", "source": "https://pika.art/blog" }, { "name": "Pika Speech", "org": "Pika", "country": "USA", "date": "2026-08-18", "precision": "day", "category": "audio", "open_weights": false, "params": "3B", "note": "Flow-matching TTS in Pika's new audio model family", "source": "https://pika.art/blog" }, { "name": "S1", "org": "Skild AI", "country": "USA", "date": "2026-08-18", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Robot model learning new tasks in-context from a single video", "source": "https://www.skild.ai/blogs/s1" }, { "name": "GEN-1.5", "org": "Generalist AI", "country": "USA", "date": "2026-08-19", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "In-context learns tasks from 12 seconds of demonstration", "source": "https://generalistai.com/blog/gen-1.5" }, { "name": "LFM2.5-DSpark", "org": "Liquid AI", "country": "USA", "date": "2026-08-20", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Up to 3.2x faster inference variant, H100 to MacBook", "source": "https://www.liquid.ai/blog/lfm2.5-dspark" }, { "name": "DeepSeek-V4-Flash-Vision-Exp", "org": "DeepSeek", "country": "China", "date": "2026-08-21", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Experimental vision model near Opus 4.8 on agent benchmarks.", "source": "https://api-docs.deepseek.com/news/news260821" }, { "name": "Granite 4.2", "org": "IBM", "country": "USA", "date": "2026-08-25", "precision": "day", "category": "language", "open_weights": true, "params": "3B / 8B / 30B", "note": "Apache-2.0 open family with reasoning effort levels for local use.", "source": "https://arstechnica.com/ai/2026/08/ibms-new-granite-4-2-models-ride-the-wave-of-interest-in-local-llms/" }, { "name": "Granite Speech 5.0", "org": "IBM", "country": "USA", "date": "2026-08-25", "precision": "day", "category": "audio", "open_weights": true, "params": "470M", "note": "Compact CTC speech recognition model", "source": "https://huggingface.co/ibm-granite/granite-speech-5.0-470m-turboctc" }, { "name": "Gemini 3.5 Transcribe", "org": "Google DeepMind", "country": "UK", "date": "2026-08-26", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-to-text in 85 languages, plus a real-time Live variant.", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5-transcribe/" }, { "name": "GLM-5.3-Flash", "org": "Zhipu AI", "country": "China", "date": "2026-08-26", "precision": "day", "category": "multimodal", "open_weights": true, "params": "320B (18B active)", "note": "First native-multimodal GLM-5 model, trained on domestic chips.", "source": "https://docs.z.ai/release-notes/new-released" }, { "name": "Qwen3.8-Flash", "org": "Alibaba", "country": "China", "date": "2026-08-26", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Low-cost hosted Qwen3.8 model with 1M context", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Qwen3.8-Flash-Next", "org": "Alibaba", "country": "China", "date": "2026-08-26", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Open, cost-optimized Flash tier of the Qwen 3.8 family.", "source": "https://the-decoder.com/alibaba-releases-qwen3-8-flash-next-targeting-ultimate-cost-efficiency/" }, { "name": "Cohere Parse", "org": "Cohere", "country": "Canada", "date": "2026-08-27", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Multimodal document parsing model producing structured Markdown", "source": "https://cohere.com/blog" }, { "name": "Gemini Omni 1.1 Flash", "org": "Google DeepMind", "country": "UK", "date": "2026-08-27", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Video model update adding scene extension, keyframes and 4K upscaling.", "source": "https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-omni-1-1-flash/" }, { "name": "Hy4 preview", "org": "Tencent", "country": "China", "date": "2026-08-28", "precision": "day", "category": "language", "open_weights": true, "params": "770B (49B active)", "note": "Tencent's open 770B flagship preview with 1M+ context.", "source": "https://www.forbes.com/sites/jonmarkman/2026/08/31/tencent-open-sources-hy4-its-770-billion-parameter-flagship-model/" }, { "name": "Isaac 0.5", "org": "Perceptron", "country": "USA", "date": "2026-08-28", "precision": "day", "category": "robotics", "open_weights": true, "params": "36B", "note": "Open embodied model beating pi0.5 and GR00T on LIBERO.", "source": "https://finance.yahoo.com/technology/ai/articles/perceptron-ai-launches-isaac-0-150000610.html" }, { "name": "Ling-3.0-flash-Fin", "org": "Ant Group", "country": "China", "date": "2026-08-28", "precision": "day", "category": "language", "open_weights": true, "params": "124B", "note": "Finance-tuned Ling model; weights open-sourced September 8.", "source": "https://technode.com/2026/08/28/ant-group-launches-finance-tuned-ling-model-plans-to-open-source-it-next-week/" }, { "name": "Mercury 2.5", "org": "Inception", "country": "USA", "date": "2026-08-31", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Diffusion LLM claiming about 1,100 tokens per second.", "source": "https://www.digitalapplied.com/blog/ai-model-releases-september-2026-tracker" }, { "name": "TimesFM-3", "org": "Google", "country": "USA", "date": "2026-08-31", "precision": "day", "category": "science", "open_weights": true, "params": "330M", "note": "Zero-shot multivariate time-series forecasting foundation model", "source": "https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/" }, { "name": "AliceAI Foundation 80B-A3B", "org": "Yandex", "country": "Russia", "date": "2026-09-01", "precision": "month", "category": "language", "open_weights": true, "params": "80B (3B active)", "note": "Yandex's open-weight hybrid MoE foundation model, 262k context", "source": "https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base" }, { "name": "Atlas", "org": "World Labs", "country": "USA", "date": "2026-09-01", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Omni world model generating and reconstructing 3D worlds from photos.", "source": "https://the-decoder.com/world-labs-unveils-atlas-a-single-ai-model-that-generates-reconstructs-and-simulates-3d-worlds-from-just-a-few-photos/" }, { "name": "Claude Fable 5.1", "org": "Anthropic", "country": "USA", "date": "2026-09-01", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Anthropic's top general-release model; 75% cheaper cache reads.", "source": "https://www.anthropic.com/claude-fable-and-mythos-5-1" }, { "name": "Claude Mythos 5.1", "org": "Anthropic", "country": "USA", "date": "2026-09-01", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Same model as Fable 5.1 with fewer safeguards; trusted access only.", "source": "https://www.anthropic.com/claude-fable-and-mythos-5-1" }, { "name": "Intern-S2", "org": "Shanghai AI Lab", "country": "China", "date": "2026-09-01", "precision": "month", "category": "science", "open_weights": true, "params": "397B", "note": "Most capable Intern model for science and long-horizon agents", "source": "https://huggingface.co/internlm/Intern-S2-397B" }, { "name": "InternLumina-U2", "org": "Shanghai AI Lab", "country": "China", "date": "2026-09-01", "precision": "day", "category": "multimodal", "open_weights": true, "params": "16B (1B active)", "note": "Diffusion LLM for omni-visual understanding, generation and editing", "source": "https://huggingface.co/internlm/InternLumina-U2" }, { "name": "Muse Voice Transcribe", "org": "Meta", "country": "USA", "date": "2026-09-01", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "MSL's first real-time streaming ASR with diarization", "source": "https://www.marktechpost.com/2026/09/01/meta-superintelligence-labs-releases-muse-voice-transcribe-one-real-time-model-for-streaming-asr-diarization-and-endpointing/" }, { "name": "Qwen-Drive 1.0", "org": "Alibaba", "country": "China", "date": "2026-09-01", "precision": "month", "category": "robotics", "open_weights": true, "params": "4B", "note": "Open driving model combining perception, traffic QA and planning.", "source": "https://the-decoder.com/qwen-drive-1-0-tells-you-why-it-brakes-just-dont-expect-the-explanation-to-match-the-maneuver/" }, { "name": "Runway Solaris", "org": "Runway", "country": "USA", "date": "2026-09-01", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "World model that generates interactive UIs frame by frame", "source": "https://runway.com/research" }, { "name": "Solaris", "org": "Runway", "country": "USA", "date": "2026-09-01", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Interface world model rendering software UIs frame by frame.", "source": "https://the-decoder.com/runways-solaris-is-an-ai-system-that-generates-software-interfaces-in-real-time/" }, { "name": "Spark X2.5 edge (4B/1.7B)", "org": "iFlytek", "country": "China", "date": "2026-09-01", "precision": "day", "category": "language", "open_weights": true, "params": "4B", "note": "First edge models with native 1M-token context", "source": "https://news.mydrivers.com/1/1147/1147820.htm" }, { "name": "YuE2", "org": "HKUST / M-A-P", "country": "China", "date": "2026-09-01", "precision": "month", "category": "audio", "open_weights": true, "params": "3B", "note": "Second-generation open full-song generation model", "source": "https://huggingface.co/m-a-p/YuE2-3B" }, { "name": "Gemini 3.8 Flash", "org": "Google DeepMind", "country": "UK", "date": "2026-09-02", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Third Gemini Flash release in six weeks.", "source": "https://9to5google.com/2026/09/02/gemini-3-8-flash-launch/" }, { "name": "Gemini 3.8 Flash Cyber", "org": "Google DeepMind", "country": "UK", "date": "2026-09-02", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Cyber-defense variant with looser mitigations, gated via Fairwind.", "source": "https://gigazine.net/gsc_news/en/20260903-gemini-3-8-flash-cyber/" }, { "name": "Muse Spark 1.3", "org": "Meta", "country": "USA", "date": "2026-09-02", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Meta closes in on frontier coding and agent performance.", "source": "https://the-decoder.com/meta-closes-in-on-the-top-with-muse-spark-1-3-and-undercuts-rivals-on-price/" }, { "name": "GPT-6 Astra", "org": "OpenAI", "country": "USA", "date": "2026-09-03", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "First model OpenAI rated Critical for cyber; touted as AGI-era step.", "source": "https://en.wikipedia.org/wiki/GPT-6_Astra" }, { "name": "GPT-6 Pro", "org": "OpenAI", "country": "USA", "date": "2026-09-03", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Astra-powered Pro tier rolled out alongside GPT-6 Astra.", "source": "https://www.forbes.com/sites/rachelwells/2026/09/06/i-tried-openais-astra-powered-gpt-6-pro-for-job-search-what-happened/" }, { "name": "GWM Worlds 2", "org": "Runway", "country": "USA", "date": "2026-09-03", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Real-time interactive 720p world simulation with audio", "source": "https://runway.com/research/introducing-gwm-worlds-2" }, { "name": "K2 Horizon", "org": "MBZUAI", "country": "UAE", "date": "2026-09-03", "precision": "day", "category": "language", "open_weights": true, "params": "0.9B-375B", "note": "Six fully open models spanning 0.9B to 375B MoE", "source": "https://en.wikipedia.org/wiki/Mohamed_bin_Zayed_University_of_Artificial_Intelligence" }, { "name": "MAI-Transcribe-2", "org": "Microsoft", "country": "USA", "date": "2026-09-03", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Claims top FLEURS accuracy across 60 languages", "source": "https://microsoft.ai/news/mai-transcribe-2-is-the-fastest-most-accurate-and-cheapest-speech-recognition-model-in-the-world/" }, { "name": "WeatherNext 3", "org": "Google DeepMind", "country": "UK", "date": "2026-09-03", "precision": "day", "category": "science", "open_weights": false, "params": "", "note": "Global weather model learning directly from live observations.", "source": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/" }, { "name": "MiniCPM5-2B", "org": "ModelBest", "country": "China", "date": "2026-09-07", "precision": "day", "category": "language", "open_weights": true, "params": "2B", "note": "Compact 2B dense on-device model", "source": "https://github.com/OpenBMB/MiniCPM" }, { "name": "Spark X2.5", "org": "iFlytek", "country": "China", "date": "2026-09-07", "precision": "day", "category": "language", "open_weights": false, "params": "293B (30B active)", "note": "MoE flagship focused on coding and agents", "source": "https://m.sohu.com/a/1072807094_115060" }, { "name": "GPT Image 2.5 (Sunburst)", "org": "OpenAI", "country": "USA", "date": "2026-09-08", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "ChatGPT Images 2.5 flagship generation and editing model", "source": "https://openai.com/index/introducing-chatgpt-images-2-5/" }, { "name": "GPT Image 2.5 Flare", "org": "OpenAI", "country": "USA", "date": "2026-09-08", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Speed-focused sibling of Sunburst at the same price.", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "GPT Image 2.5 Sunburst", "org": "OpenAI", "country": "USA", "date": "2026-09-08", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Precision-focused image model behind ChatGPT Images 2.5.", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "AuK", "org": "Tencent", "country": "China", "date": "2026-09-09", "precision": "day", "category": "audio", "open_weights": true, "params": "", "note": "Open foundation model for speech generation and editing", "source": "https://github.com/Tencent-Hunyuan/AuK" }, { "name": "Suno v6", "org": "Suno", "country": "USA", "date": "2026-09-09", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Suno's first models built on label-licensed music; sued again.", "source": "https://the-decoder.com/suno-launches-v6-music-models-built-with-warner-bmg-and-believe/" }, { "name": "Suno v6-mini", "org": "Suno", "country": "USA", "date": "2026-09-09", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Fast free-tier v6 model", "source": "https://en.wikipedia.org/wiki/Suno_AI" }, { "name": "Suno v6-wild", "org": "Suno", "country": "USA", "date": "2026-09-09", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Experimental, less predictable v6 variant", "source": "https://en.wikipedia.org/wiki/Suno_AI" }, { "name": "DeepSeek-V4.1-Flash", "org": "DeepSeek", "country": "China", "date": "2026-09-10", "precision": "day", "category": "language", "open_weights": true, "params": "552B", "note": "New causal encoder-decoder architecture with native multimodality.", "source": "https://api-docs.deepseek.com/news/news260910" }, { "name": "North Small Translate", "org": "Cohere", "country": "Canada", "date": "2026-09-10", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Open-weight MoE machine translation model for 50+ languages", "source": "https://cohere.com/blog" }, { "name": "SWE-2", "org": "Cognition", "country": "USA", "date": "2026-09-10", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Kimi K3 post-trained coder near Fable 5.1 at 64% lower cost.", "source": "https://www.marktechpost.com/2026/09/12/cognition-releases-swe-2-a-kimi-k3-post-trained-coding-model-that-matches-fable-5-1-on-frontiercode-at-64-lower-cost/" }, { "name": "Atria Dawn Preview", "org": "ATRIA", "country": "China", "date": "2026-09-11", "precision": "day", "category": "agent", "open_weights": true, "params": "", "note": "Open model for long-horizon scientific research agents.", "source": "https://llm-stats.com/llm-updates" }, { "name": "Fugu Max", "org": "Sakana AI", "country": "Japan", "date": "2026-09-11", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Lower-cost Fugu orchestration tier at $2/$6 per million tokens.", "source": "https://gigazine.net/gsc_news/en/20260914-fugu-ultra-v2/" }, { "name": "Fugu Max and Fugu Ultra v2", "org": "Sakana AI", "country": "Japan", "date": "2026-09-11", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "New Fugu tiers targeting the cost-performance Pareto frontier", "source": "https://sakana.ai/blog/" }, { "name": "Fugu Ultra v2", "org": "Sakana AI", "country": "Japan", "date": "2026-09-11", "precision": "day", "category": "agent", "open_weights": false, "params": "", "note": "Orchestration system Sakana claims beats GPT-6 Astra.", "source": "https://gigazine.net/gsc_news/en/20260914-fugu-ultra-v2/" }, { "name": "Kimi K2.8 Preview", "org": "Moonshot AI", "country": "China", "date": "2026-09-11", "precision": "day", "category": "code", "open_weights": false, "params": "", "note": "Coding and agent preview model with 1M context.", "source": "https://llm-stats.com/ai-news" }, { "name": "Ling-3.0-flash-VL", "org": "Ant Group", "country": "China", "date": "2026-09-11", "precision": "day", "category": "multimodal", "open_weights": true, "params": "124B (5.5B active)", "note": "MIT-licensed vision-language member of the Ling 3.0 family.", "source": "https://cryptobriefing.com/ant-group-ling-3-flash-vl-intelligence-index/" }, { "name": "Scribe v2 Medical", "org": "ElevenLabs", "country": "USA", "date": "2026-09-11", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Clinical-audio speech recognition with 35% fewer errors", "source": "https://elevenlabs.io/docs/changelog" }, { "name": "Eleven Music v2.5", "org": "ElevenLabs", "country": "USA", "date": "2026-09-13", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Fuller, more natural AI music generation via app and API.", "source": "https://the-decoder.com/elevenlabs-makes-music-v2-5-available-via-app-and-api-with-free-and-pro-tier-options/" }, { "name": "Iris-mini", "org": "AllSpark", "country": "China", "date": "2026-09-13", "precision": "day", "category": "agent", "open_weights": true, "params": "35B", "note": "Small open-weight deep-search agent model.", "source": "https://the-decoder.com/iris-mini-and-iris-pro-are-the-strongest-open-weight-search-agents-in-their-class/" }, { "name": "Iris-pro", "org": "AllSpark", "country": "China", "date": "2026-09-13", "precision": "day", "category": "agent", "open_weights": true, "params": "397B", "note": "Strongest open-weight search agent in its class, Qwen-based.", "source": "https://the-decoder.com/iris-mini-and-iris-pro-are-the-strongest-open-weight-search-agents-in-their-class/" }, { "name": "Intern-W0", "org": "Shanghai AI Lab", "country": "China", "date": "2026-09-14", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Physical world model for force-tactile robotics", "source": "https://pandaily.com/shanghai-ai-lab-intern-w0-physical-world-model" }, { "name": "Gemini 3.8 Live", "org": "Google DeepMind", "country": "UK", "date": "2026-09-15", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Native voice-agent model in 97 languages, rivaling GPT-Live-1.", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/" }, { "name": "Gemini 3.8 Live Extended Thinking", "org": "Google DeepMind", "country": "UK", "date": "2026-09-15", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Reasoning variant of the Live voice model for complex tasks.", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/" }, { "name": "Periodic Neon", "org": "Periodic Labs", "country": "USA", "date": "2026-09-15", "precision": "day", "category": "science", "open_weights": false, "params": "1T", "note": "Materials-science LLM trained with lab-in-the-loop learning", "source": "https://periodic.com/news/nature-is-our-learning-environment" }, { "name": "Vidu S2", "org": "ShengShu", "country": "China", "date": "2026-09-16", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Second Vidu S-series video model", "source": "https://www.prnewswire.com/news/shengshu-technology/" }, { "name": "Grok Voice Transcribe 2.0", "org": "xAI", "country": "USA", "date": "2026-09-17", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Speech-to-text model claimed twice as accurate as Transcribe 1.0.", "source": "https://x.ai/news/grok-voice-transcribe-2" }, { "name": "Helix 2.5", "org": "Figure", "country": "USA", "date": "2026-09-17", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Zero-shot household tasks in 30 unseen homes via Index pretraining", "source": "https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization" }, { "name": "Qwen3.8-Omni-Flash", "org": "Alibaba", "country": "China", "date": "2026-09-17", "precision": "day", "category": "multimodal", "open_weights": false, "params": "", "note": "Agentic omni model: text, image, audio, video input; 1M context", "source": "https://www.alibabacloud.com/help/en/model-studio/newly-released-models" }, { "name": "Ternary Bonsai 2 27B", "org": "PrismML", "country": "USA", "date": "2026-09-17", "precision": "day", "category": "language", "open_weights": true, "params": "27B", "note": "Ternary-weight Qwen3.8-27B squeezed from 53.8 GB to 5.9 GB.", "source": "https://aiweekly.co/alerts/prismml-compresses-qwen38-27b-to-593-gb-with-ternary-weights" }, { "name": "HiDream-O1-Video-1.0", "org": "HiDream.ai", "country": "China", "date": "2026-09-18", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Native omnimodal video model focused on physical consistency", "source": "https://www.financialcontent.com/article/getnews-2026-9-18-hidream-unveils-hidream-o1-video-10-a-native-omnimodal-video-model-built-for-physical-consistency" }, { "name": "Step 5 Preview", "org": "StepFun", "country": "China", "date": "2026-09-18", "precision": "day", "category": "language", "open_weights": false, "params": "600B (27B active)", "note": "Flagship agentic MoE with 1M context; open weights promised Oct 15", "source": "https://www.marktechpost.com/2026/09/20/stepfun-launches-step-5-preview/" }, { "name": "Gander", "org": "Tencent", "country": "China", "date": "2026-09-20", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Voice assistant model that keeps talking while agents work.", "source": "https://the-decoder.com/tencents-gander-aims-to-keep-talking-while-it-works-in-the-background/" }, { "name": "Qwen-Image-2.1", "org": "Alibaba", "country": "China", "date": "2026-09-20", "precision": "day", "category": "image", "open_weights": true, "params": "7B", "note": "Open 7B image model, now under a non-commercial license.", "source": "https://the-decoder.com/alibabas-open-weight-qwen-image-2-1-claims-to-beat-closed-models-in-image-generation-with-just-7-billion-parameters/" }, { "name": "Grok 4.7", "org": "xAI", "country": "USA", "date": "2026-09-21", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Larger base model and longer RL run; 500K context.", "source": "https://github.blog/changelog/2026-09-21-grok-4-7-is-now-available-in-github-copilot/" }, { "name": "Hy Image 3.5 Preview", "org": "Tencent", "country": "China", "date": "2026-09-21", "precision": "day", "category": "image", "open_weights": false, "params": "", "note": "Next Hunyuan image model aimed at ByteDance and Alibaba rivals", "source": "https://www.briefs.co/news/tencent-debuts-hy-image-3-5-preview-amid-alibaba-ai-event/" }, { "name": "MiMo-V2.6-Flash", "org": "Xiaomi", "country": "China", "date": "2026-09-21", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Near-Pro performance at a third of the API price", "source": "https://news.mydrivers.com/1/1152/1152935.htm" }, { "name": "MiMo-V2.6-Pro", "org": "Xiaomi", "country": "China", "date": "2026-09-21", "precision": "day", "category": "multimodal", "open_weights": true, "params": "", "note": "Top open-weights model on AA index; RL training livestreamed", "source": "https://news.mydrivers.com/1/1152/1152935.htm" }, { "name": "Claude Opus 5.5", "org": "Anthropic", "country": "USA", "date": "2026-09-22", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "Fable-5.1-level work at 40% lower cost than Opus 5.", "source": "https://www.anthropic.com/claude-opus-5-5" }, { "name": "Gemini 3.8 Flash TTS", "org": "Google DeepMind", "country": "UK", "date": "2026-09-22", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Expressive TTS with voice design and consented voice replication", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/" }, { "name": "Gemini 3.8 Flash-Lite TTS", "org": "Google DeepMind", "country": "UK", "date": "2026-09-22", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Low-cost, high-volume text-to-speech variant", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/" }, { "name": "GPT-6 Luna", "org": "OpenAI", "country": "USA", "date": "2026-09-22", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Smallest, cheapest GPT-6 reasoning tier.", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "GPT-6 Sol", "org": "OpenAI", "country": "USA", "date": "2026-09-22", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Cheaper GPT-6 reasoning tier below Astra; halved flagship prices.", "source": "https://developers.openai.com/api/docs/changelog" }, { "name": "Ming-Image-0.1-Design", "org": "Ant Group", "country": "China", "date": "2026-09-22", "precision": "day", "category": "image", "open_weights": true, "params": "6B", "note": "Open design-focused image models with layer decomposition.", "source": "https://news.cnyes.com/news/id/6613828" }, { "name": "Solar Mini 4", "org": "Upstage", "country": "South Korea", "date": "2026-09-22", "precision": "day", "category": "language", "open_weights": false, "params": "35B (3B active)", "note": "Compact agentic model from Upstage.", "source": "https://www.digitalapplied.com/blog/ai-model-releases-september-2026-tracker" }, { "name": "FLUX 3 Action", "org": "Black Forest Labs", "country": "Germany", "date": "2026-09-23", "precision": "day", "category": "robotics", "open_weights": true, "params": "7B", "note": "Image lab's open world-action robot model tops RoboLab-120.", "source": "https://www.marktechpost.com/2026/09/24/black-forest-labs-releases-flux-3-action-a-7b-open-weights-world-action-model-that-tops-robolab-120/" }, { "name": "Muse Realtime Avatar", "org": "Meta", "country": "USA", "date": "2026-09-23", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Real-time talking avatars for Muse agent, shown at Connect", "source": "https://techcrunch.com/2026/09/23/everything-new-coming-to-metas-ai-agent-muse/" }, { "name": "Qwen Audio 3.1", "org": "Alibaba", "country": "China", "date": "2026-09-23", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Five ASR, TTS and realtime models; prices cut up to 95%.", "source": "https://the-decoder.com/alibaba-launches-qwen-audio-3-1-with-five-new-models-and-slashes-ai-audio-prices-by-up-to-95-percent/" }, { "name": "Ember-1", "org": "Fireworks AI", "country": "USA", "date": "2026-09-24", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Kimi K3 post-train using about 40% fewer reasoning tokens.", "source": "https://www.marktechpost.com/2026/09/28/fireworks-ai-releases-ember-1-a-post-trained-kimi-k3-that-uses-about-40-fewer-tokens/" }, { "name": "Gemini 3.8 Live with Live Avatar", "org": "Google DeepMind", "country": "UK", "date": "2026-09-24", "precision": "day", "category": "video", "open_weights": false, "params": "", "note": "Real-time talking avatar with lip-sync on the live dialogue model.", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-with-live-avatar/" }, { "name": "LFM2.5-VL-DSpark", "org": "Liquid AI", "country": "USA", "date": "2026-09-24", "precision": "day", "category": "language", "open_weights": true, "params": "", "note": "Accelerated vision-language variant for edge deployment", "source": "https://www.liquid.ai/blog/lfm2-5-vl-dspark" }, { "name": "Sarvam Vision 2.1", "org": "Sarvam AI", "country": "India", "date": "2026-09-24", "precision": "day", "category": "vision", "open_weights": false, "params": "", "note": "Indic document intelligence with handwritten-text recognition.", "source": "https://www.sarvam.ai/blogs" }, { "name": "Perceptron Mk1.5", "org": "Perceptron", "country": "USA", "date": "2026-09-25", "precision": "day", "category": "robotics", "open_weights": false, "params": "", "note": "Embodied agent controller with multimodal input, structured output.", "source": "https://www.digitalapplied.com/blog/ai-model-releases-september-2026-tracker" }, { "name": "LongCat-2.5-Preview", "org": "Meituan", "country": "China", "date": "2026-09-26", "precision": "day", "category": "multimodal", "open_weights": false, "params": "1.6T (MoE)", "note": "Natively multimodal long-horizon agent model with 1M context", "source": "https://t.cj.sina.cn/articles/view/1826017320/6cd6d02802001x8lu" }, { "name": "Nemotron 3 Diarization", "org": "NVIDIA", "country": "USA", "date": "2026-09-27", "precision": "day", "category": "audio", "open_weights": true, "params": "100M", "note": "Free real-time diarization identifying up to eight speakers.", "source": "https://the-decoder.com/nvidia-drops-a-free-100m-parameter-model-that-identifies-up-to-eight-speakers-in-real-time/" }, { "name": "Claude Sonnet 5.5", "org": "Anthropic", "country": "USA", "date": "2026-09-28", "precision": "day", "category": "language", "open_weights": false, "params": "", "note": "30% faster than Sonnet 5 and up to 30% cheaper per task.", "source": "https://www.anthropic.com/claude-sonnet-5-5" }, { "name": "Eleven v4", "org": "ElevenLabs", "country": "USA", "date": "2026-09-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Most expressive ElevenLabs TTS, with low-latency v4 Turbo variant.", "source": "https://elevenlabs.io/blog/eleven-v4" }, { "name": "Eleven v4 Turbo", "org": "ElevenLabs", "country": "USA", "date": "2026-09-28", "precision": "day", "category": "audio", "open_weights": false, "params": "", "note": "Real-time v4 variant with ~100 ms latency", "source": "https://elevenlabs.io/blog/eleven-v4" }, { "name": "Holo4", "org": "H Company", "country": "France", "date": "2026-09-28", "precision": "day", "category": "agent", "open_weights": true, "params": "27B / 35B (3B active)", "note": "85.2% on OSWorld at $0.08 per task", "source": "https://hcompany.ai/newsroom/holo4" }, { "name": "GPT-6.1 Sol", "org": "OpenAI", "country": "USA", "date": "2026-09-29", "precision": "day", "category": "reasoning", "open_weights": false, "params": "", "note": "Near-Astra performance at one-fifth of Astra's token price.", "source": "https://openai.com/index/introducing-gpt-6-1-sol/" } ] }