AI/SI Model Registry · The AI / SI Encyclopedia

A living register of 55 historically or technically significant AI models and systems, from the 1958 perceptron to systems announced in September and October 2026. Each record states developer, source date, architecture family where the developer disclosed it, modalities, access or licence, a primary source, limitations stated by the source, and status. It is a record of what sources say, not a ranking: it contains no benchmark leaderboards and no claim that listed systems are the most capable of their era.

#Conventions and method

Source-first. Every record rests on a primary source (paper, model card, or developer announcement) that was opened or queried on the confirmation date given in the record. Titles, author counts, and dates were copied from the arXiv abstract page, Crossref, PubMed, the publisher page, or the developer page, not from memory.
"undisclosed" means the primary source was read and the developer does not state the detail. "not extracted" means the detail may exist in the full text but was not retrieved in this drafting pass; it is not a claim that it is undisclosed. Nothing is inferred.
Date is the date on the primary source (arXiv v1 date, journal online date, or announcement date). It is not necessarily the date of first public availability.
Developer gives the organisation as stated on the developer page or in the author affiliations of the paper where those were retrieved (DeepMind papers, OpenAI GPT-1/GPT-2 PDFs, Anthropic, Google, Meta, xAI pages). For other arXiv entries the organisation is the one the encyclopedia associates with the first author and the affiliation line was not separately re-extracted; the first author and author count are given so the reader can check.
Licence entries distinguish a code-repository licence (read from the GitHub API) from a weights licence (read from the Hugging Face model card). A code licence is not a weights licence. Where neither was checked the entry says "not checked".
2025-2026 systems carry a confirmation date. Only systems whose developer page (or a developer-run system-card site) was fetched on 2026-10-15 are included. Vendor-reported figures (prices, benchmark scores, context lengths) are deliberately omitted because they change; see the Living Reference collections.
Excluded or unconfirmed items are listed at the end.
Statuses: Historical milestone, Superseded by developer's later releases, Listed on developer page as of 2026-10-15, Open weights, Not released (preview). "Current availability not checked" is stated where it was not verified.

#Records

#M-01 Perceptron

Developer: F. Rosenblatt (affiliation not extracted)
Date: 1958
Family / architecture: Perceptron; the paper's title describes "a probabilistic model for information storage and organization in the brain". Further architectural detail not extracted.
Modalities: not extracted
Access / licence: not applicable (published model); not checked
Primary source: Rosenblatt, F. (1958). The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 65, 386-408. https://doi.org/10.1037/h0042519 (Crossref, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Historical milestone

#M-02 ELIZA

Developer: J. Weizenbaum (affiliation not extracted)
Date: 1966
Family / architecture: natural-language communication program (title: "A Computer Program For the Study of Natural Language Communication Between Man and Machine"); internals not extracted
Modalities: text
Access / licence: not checked
Primary source: Weizenbaum, J. (1966). ELIZA - A Computer Program For the Study of Natural Language Communication Between Man and Machine. Communications of the ACM, 9, 36-45. https://doi.org/10.1145/365153.365168 (Crossref)
Limitations stated by source: not extracted
Status: Historical milestone

#M-03 Long Short-Term Memory (LSTM)

Developer: S. Hochreiter, J. Schmidhuber (Technische Universität München, per PubMed author information)
Date: 1997 (issue dated 1997-11-15)
Family / architecture: recurrent network with gate units and constant error carousels, trained by a gradient-based method; abstract reports bridging time lags beyond 1000 discrete steps on artificial tasks
Modalities: sequence data (experiments in the paper use artificial data)
Access / licence: not applicable; not checked
Primary source: Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation, 9(8), 1735-1780. https://doi.org/10.1162/neco.1997.9.8.1735 (PubMed record, confirmed 2026-10-15)
Limitations stated by source: none stated in the abstract; full text not extracted
Status: Historical milestone

#M-04 Deep Blue

Developer: M. Campbell, A. J. Hoane, F. Hsu (3 authors; affiliations not extracted)
Date: description paper 2002-01 (Artificial Intelligence 134, 57-83); the system is the chess machine discussed in that paper
Family / architecture: special-purpose chess system; details not extracted
Modalities: chess positions
Access / licence: not applicable
Primary source: Campbell, M., Hoane, A. J., & Hsu, F. (2002). Deep Blue. Artificial Intelligence, 134, 57-83. https://doi.org/10.1016/S0004-3702(01)00129-1 (Crossref)
Limitations stated by source: not extracted
Status: Historical milestone. See HIST-009 for the match history.

#M-05 LeNet-5 (gradient-based document recognition)

Developer: Y. LeCun, L. Bottou, Y. Bengio, P. Haffner (4 authors; affiliations not extracted)
Date: 1998
Family / architecture: convolutional neural network trained by gradient-based learning (title: "Gradient-based learning applied to document recognition"); the paper's name for the specific network is not extracted
Modalities: images of characters/documents
Access / licence: not applicable
Primary source: LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86, 2278-2324. https://doi.org/10.1109/5.726791 (Crossref)
Limitations stated by source: not extracted
Status: Historical milestone

#M-06 AlexNet

Developer: A. Krizhevsky, I. Sutskever, G. E. Hinton
Date: 2012
Family / architecture: deep convolutional neural network for ImageNet classification (per title)
Modalities: images
Access / licence: not checked
Primary source: Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems 25 (NIPS 2012). https://papers.nips.cc/paper_files/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html (proceedings page, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Historical milestone. See HIST-005.

#M-07 word2vec

Developer: T. Mikolov, K. Chen, G. Corrado, J. Dean (4 authors)
Date: 2013-01-16 (arXiv v1)
Family / architecture: word-embedding estimation methods ("Efficient Estimation of Word Representations in Vector Space"); details not extracted
Modalities: text
Access / licence: not checked
Primary source: Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient Estimation of Word Representations in Vector Space. arXiv:1301.3781 (arXiv abstract page, confirmed 2026-10-15). Venue not confirmed.
Limitations stated by source: not extracted
Status: Historical milestone

#M-08 Generative Adversarial Networks (GAN)

Developer: I. J. Goodfellow et al. (8 authors)
Date: 2014-06-10 (arXiv v1)
Family / architecture: adversarial framework training a generative model G and a discriminative model D as a minimax game; abstract notes G and D may be multilayer perceptrons
Modalities: data-distribution agnostic (framework)
Access / licence: not applicable
Primary source: Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., et al. (2014). Generative Adversarial Networks. arXiv:1406.2661 (arXiv abstract page, confirmed 2026-10-15). Venue not confirmed.
Limitations stated by source: not extracted
Status: Historical milestone

#M-09 Deep Q-Network (DQN)

Developer: V. Mnih et al., Google DeepMind (19 authors; affiliation per PubMed)
Date: 2015-02-26 (issue date per PubMed; Crossref issued date 2015-02-25)
Family / architecture: deep reinforcement learning agent (title: "Human-level control through deep reinforcement learning"); the abstract excerpt retrieved frames the work as learning from high-dimensional sensory input
Modalities: visual game input (per abstract framing; Atari detail not extracted)
Access / licence: not checked
Primary source: Mnih, V., Kavukcuoglu, K., Silver, D., et al. (2015). Human-level control through deep reinforcement learning. Nature, 518, 529-533. https://doi.org/10.1038/nature14236 (PubMed, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Historical milestone

#M-10 ResNet

Developer: K. He, X. Zhang, S. Ren, J. Sun (4 authors)
Date: 2015-12-10 (arXiv v1)
Family / architecture: residual convolutional networks; abstract reports evaluation with depth up to 152 layers on ImageNet
Modalities: images
Access / licence: not checked
Primary source: He, K., Zhang, X., Ren, S., & Sun, J. (2015). Deep Residual Learning for Image Recognition. arXiv:1512.03385 (arXiv, confirmed 2026-10-15). Venue not confirmed.
Limitations stated by source: abstract: deeper networks are more difficult to train (the motivation for the method)
Status: Historical milestone

#M-11 AlphaGo

Developer: D. Silver et al., Google DeepMind (20 authors)
Date: 2016-01-27 (online, per Crossref)
Family / architecture: value networks and policy networks combined with Monte Carlo tree search; trained by supervised learning from human expert games and reinforcement learning from self-play (abstract)
Modalities: Go board positions
Access / licence: not released (paper only; not checked further)
Primary source: Silver, D., Huang, A., Maddison, C. J., et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529, 484-489. https://doi.org/10.1038/nature16961 (PubMed, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Historical milestone. See HIST-009.

#M-12 Transformer

Developer: A. Vaswani et al. (8 authors)
Date: 2017-06-12 (arXiv v1)
Family / architecture: encoder-decoder network based solely on attention, with no recurrence or convolution; abstract reports 28.4 BLEU on WMT 2014 English-to-German (a 2017 figure, dated)
Modalities: text (machine translation experiments)
Access / licence: not applicable
Primary source: Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention Is All You Need. arXiv:1706.03762 (arXiv, confirmed 2026-10-15). See TECH-ARCH-001.
Limitations stated by source: none stated in the abstract
Status: Historical milestone; architecture family underlying many later entries

#M-13 AlphaGo Zero

Developer: D. Silver et al., DeepMind (17 authors)
Date: 2017-10-18
Family / architecture: reinforcement learning from self-play only, with no human data or domain knowledge beyond the game rules; one network predicts move selections and the winner; abstract reports a 100-0 result against the previously published AlphaGo
Modalities: Go board positions
Access / licence: not released (paper only; not checked further)
Primary source: Silver, D., Schrittwieser, J., Simonyan, K., et al. (2017). Mastering the game of Go without human knowledge. Nature, 550, 354-359. https://doi.org/10.1038/nature24270 (PubMed, confirmed 2026-10-15)
Limitations stated by source: none stated in the abstract
Status: Superseded by AlphaZero (M-16) in the developer's own line

#M-14 GPT-1 (generative pre-training)

Developer: A. Radford, K. Narasimhan, T. Salimans, I. Sutskever (OpenAI, per paper header)
Date: 2018 (PDF carries no day; announcement-page date not retrievable because openai.com returned HTTP 403)
Family / architecture: generative pre-training of a language model on unlabeled text followed by discriminative fine-tuning per task; abstract reports gains over state of the art on 9 of 12 tasks studied
Modalities: text
Access / licence: not checked
Primary source: Radford, A., Narasimhan, K., Salimans, T., & Sutskever, I. (2018). Improving Language Understanding by Generative Pre-Training. OpenAI. https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf (PDF opened 2026-10-15). The paper's title and author list were read from the PDF's first page.
Limitations stated by source: not extracted
Status: Historical milestone

#M-15 BERT

Developer: J. Devlin, M.-W. Chang, K. Lee, K. Toutanova (4 authors)
Date: 2018-10-11 (arXiv v1)
Family / architecture: Bidirectional Encoder Representations from Transformers; pre-trains deep bidirectional representations from unlabeled text, fine-tuned with one extra output layer (abstract)
Modalities: text
Access / licence: code repository google-research/bert: Apache-2.0; weights: google-bert/bert-base-uncased card: Apache-2.0 (GitHub API and Hugging Face API, 2026-10-15)
Primary source: Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 (arXiv, confirmed 2026-10-15)
Limitations stated by source: none stated in the abstract
Status: Open weights

#M-16 AlphaZero

Developer: D. Silver et al., DeepMind (13 authors)
Date: 2018-12-07 (Science)
Family / architecture: a single reinforcement-learning algorithm, generalising AlphaGo Zero, applied to chess, shogi, and Go from random play with only the rules given (abstract)
Modalities: board-game positions
Access / licence: not released (not checked further)
Primary source: Silver, D., Hubert, T., Schrittwieser, J., et al. (2018). A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362, 1140-1144. https://doi.org/10.1126/science.aar6404 (PubMed, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Historical milestone. See HIST-009.

#M-17 GPT-2

Developer: A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever (OpenAI, per paper header)
Date: 2019 (PDF carries no day; announcement date not retrievable, openai.com HTTP 403)
Family / architecture: Transformer language model; the largest model is described as 1.5B parameters, trained on WebText (millions of web pages)
Modalities: text
Access / licence: weights on Hugging Face (openai-community/gpt2) tagged MIT; code repository openai/gpt-2 licence field: NOASSERTION (GitHub API, 2026-10-15)
Primary source: Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language Models are Unsupervised Multitask Learners. OpenAI. https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf (PDF opened 2026-10-15)
Limitations stated by source: abstract states the largest model "still underfits WebText"
Status: Open weights

#M-18 AlphaStar

Developer: O. Vinyals et al., DeepMind (42 authors; one affiliation outside DeepMind)
Date: 2019-10-30 (online)
Family / architecture: multi-agent reinforcement learning (title: "Grandmaster level in StarCraft II using multi-agent reinforcement learning")
Modalities: game state of StarCraft II
Access / licence: not released (not checked further)
Primary source: Vinyals, O., Babuschkin, I., Czarnecki, W. M., et al. (2019). Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575, 350-354. https://doi.org/10.1038/s41586-019-1724-z (PubMed, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Historical milestone

#M-19 GPT-3

Developer: T. B. Brown et al. (31 authors; OpenAI affiliation not separately extracted)
Date: 2020-05-28 (arXiv v1)
Family / architecture: autoregressive language model with 175 billion parameters, applied without gradient updates or fine-tuning, tasks and few-shot demonstrations given as text (abstract)
Modalities: text
Access / licence: not stated in abstract
Primary source: Brown, T. B., Mann, B., Ryder, N., et al. (2020). Language Models are Few-Shot Learners. arXiv:2005.14165 (arXiv, confirmed 2026-10-15)
Limitations stated by source: abstract refers to current NLP systems' difficulty in learning from few examples as the motivation; model-specific limitations not extracted
Status: Superseded by developer's later releases; current availability not checked

#M-20 Vision Transformer (ViT)

Developer: A. Dosovitskiy et al. (12 authors)
Date: 2020-10-22 (arXiv v1)
Family / architecture: pure Transformer applied to sequences of image patches
Modalities: images
Access / licence: code repository google-research/vision_transformer: Apache-2.0 (GitHub API). Weights licence not checked.
Primary source: Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al. (2020). An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv:2010.11929 (arXiv, confirmed 2026-10-15)
Limitations stated by source: abstract says attains excellent results "when pre-trained on large amounts of data"
Status: Historical milestone

#M-21 AlphaFold 2

Developer: J. Jumper et al., DeepMind (34 authors)
Date: 2021-07-15 (online); the paper reports a system evaluated before this date (CASP14 date not extracted)
Family / architecture: neural network predicting protein structure ("Highly accurate protein structure prediction with AlphaFold"); internals not extracted
Modalities: protein sequences to 3D structure
Access / licence: code repository google-deepmind/alphafold: Apache-2.0 (GitHub API). Weights terms not checked.
Primary source: Jumper, J., Evans, R., Pritzel, A., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583-589. https://doi.org/10.1038/s41586-021-03819-2 (PubMed, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Superseded by AlphaFold 3 (M-38) in the developer's own line for biomolecular complexes; both documented here

#M-22 CLIP

Developer: A. Radford, J. W. Kim, C. Hallacy et al. (12 authors)
Date: 2021-02-26 (arXiv v1)
Family / architecture: image representation learning by predicting which caption goes with which image, on 400 million (image, text) pairs collected from the internet (abstract)
Modalities: image and text
Access / licence: code repository openai/CLIP: MIT (GitHub API); HF card licence field empty
Primary source: Radford, A., Kim, J. W., Hallacy, C., et al. (2021). Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020 (arXiv, confirmed 2026-10-15)
Limitations stated by source: abstract notes fixed-category supervision limits generality (the motivation)
Status: Open code; weights licence not checked

#M-23 DALL-E (original)

Developer: A. Ramesh et al. (8 authors)
Date: 2021-02-24 (arXiv v1)
Family / architecture: transformer that autoregressively models text and image tokens as a single stream of data (abstract)
Modalities: text to image
Access / licence: not checked
Primary source: Ramesh, A., Pavlov, M., Goh, G., et al. (2021). Zero-Shot Text-to-Image Generation. arXiv:2102.12092 (arXiv, confirmed 2026-10-15)
Limitations stated by source: none stated in the abstract
Status: Historical milestone

#M-24 Codex

Developer: M. Chen et al. (58 authors)
Date: 2021-07-07 (arXiv v1)
Family / architecture: a GPT language model fine-tuned on publicly available code from GitHub; a distinct production version powers GitHub Copilot (abstract)
Modalities: text and code
Access / licence: not stated in abstract; HumanEval evaluation set repository openai/human-eval: MIT (GitHub API)
Primary source: Chen, M., Tworek, J., Jun, H., et al. (2021). Evaluating Large Language Models Trained on Code. arXiv:2107.03374 (arXiv, confirmed 2026-10-15)
Limitations stated by source: abstract states difficulty with docstrings describing long chains of operations and with binding operations to variables
Status: Superseded by developer's later releases; current availability not checked

#M-25 Latent Diffusion Models / Stable Diffusion

Developer: R. Rombach et al. (5 authors) for the method; Stability AI for the public Stable Diffusion release, with Hugging Face and CoreWeave named as collaborators on the release (Stability AI announcement)
Date: method 2021-12-20 (arXiv v1); public release announcement 22 August 2022 (the Stability page text shows "August 22" without a year; the year 2022 is taken from the Hugging Face repository creation date 2022-08-20 for CompVis/stable-diffusion-v1-4)
Family / architecture: diffusion models applied in the latent space of pretrained autoencoders (abstract of the method paper)
Modalities: text to image (as released); the method paper's abstract refers to image synthesis and beyond
Access / licence: weights: Creative ML OpenRAIL-M (Stability AI page; Hugging Face tag creativeml-openrail-m). Code repository CompVis/latent-diffusion: MIT; CompVis/stable-diffusion: NOASSERTION (GitHub API).
Primary source: Rombach, R., Blattmann, A., Lorenz, D., et al. (2021). High-Resolution Image Synthesis with Latent Diffusion Models. arXiv:2112.10752 (arXiv); Stability AI, Stable Diffusion Public Release, http://stability.ai/news-updates/stable-diffusion-public-release (fetched 2026-10-15)
Limitations stated by source: abstract states pixel-space diffusion models are expensive to optimise and run (the motivation)
Status: Open weights (v1.4 listed on the Stability page)

#M-26 Chinchilla

Developer: J. Hoffmann et al. (22 authors; DeepMind affiliation not separately extracted)
Date: 2022-03-29 (arXiv v1)
Family / architecture: compute-optimal training study of transformer language models over 400+ models from 70 million to over 16 billion parameters; finding that model size and training tokens should scale equally (abstract). Chinchilla is the model trained under that rule.
Modalities: text
Access / licence: not checked
Primary source: Hoffmann, J., Borgeaud, S., Mensch, A., et al. (2022). Training Compute-Optimal Large Language Models. arXiv:2203.15556 (arXiv, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Historical milestone

#M-27 InstructGPT

Developer: L. Ouyang et al. (20 authors)
Date: 2022-03-04 (arXiv v1)
Family / architecture: GPT-3 fine-tuned with supervised learning on labeler demonstrations, then with reinforcement learning from human feedback on output rankings (abstract)
Modalities: text
Access / licence: not stated in abstract
Primary source: Ouyang, L., Wu, J., Jiang, X., et al. (2022). Training language models to follow instructions with human feedback. arXiv:2203.02155 (arXiv, confirmed 2026-10-15)
Limitations stated by source: abstract describes improvements in truthfulness (details truncated in retrieval); remaining limitations not extracted
Status: Historical milestone. See TECH-ALIGN-001.

#M-28 PaLM

Developer: A. Chowdhery et al. (67 authors; Google affiliation not separately extracted)
Date: 2022-04-05 (arXiv v1)
Family / architecture: 540-billion-parameter densely activated Transformer language model trained on 6144 TPU v4 chips using Pathways (abstract)
Modalities: text
Access / licence: not checked
Primary source: Chowdhery, A., Narang, S., Devlin, J., et al. (2022). PaLM: Scaling Language Modeling with Pathways. arXiv:2204.02311 (arXiv, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Historical milestone

#M-29 Whisper

Developer: A. Radford, J. W. Kim, T. Xu et al. (6 authors)
Date: 2022-12-06 (arXiv v1)
Family / architecture: speech-processing models trained to predict transcripts of audio from the internet, at 680,000 hours of multilingual and multitask supervision (abstract)
Modalities: audio to text
Access / licence: code repository openai/whisper: MIT; openai/whisper-large-v3 card: Apache-2.0 (GitHub and Hugging Face APIs). Abstract states models and inference code are released.
Primary source: Radford, A., Kim, J. W., Xu, T., et al. (2022). Robust Speech Recognition via Large-Scale Weak Supervision. arXiv:2212.04356 (arXiv, confirmed 2026-10-15)
Limitations stated by source: abstract: approaches human accuracy and robustness; limitations not extracted
Status: Open weights

#M-30 LLaMA

Developer: H. Touvron et al. (14 authors; Meta affiliation not separately extracted)
Date: 2023-02-27 (arXiv v1)
Family / architecture: collection of foundation language models from 7B to 65B parameters trained on publicly available datasets only (abstract)
Modalities: text
Access / licence: abstract: released "to the research community"; licence text not checked
Primary source: Touvron, H., Lavril, T., Izacard, G., et al. (2023). LLaMA: Open and Efficient Foundation Language Models. arXiv:2302.13971 (arXiv, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Superseded by developer's later releases (M-31, M-39, M-44)

#M-31 Llama 2

Developer: H. Touvron et al. (68 authors)
Date: 2023-07-18 (arXiv v1)
Family / architecture: pretrained and fine-tuned LLMs from 7B to 70B parameters; Llama 2-Chat optimised for dialogue (abstract)
Modalities: text
Access / licence: weights: Hugging Face meta-llama/Llama-2-7b-hf licence tag "llama2", access gated (HF API, 2026-10-15)
Primary source: Touvron, H., Martin, L., Stone, K., et al. (2023). Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv:2307.09288 (arXiv)
Limitations stated by source: not extracted
Status: Superseded by developer's later releases; weights gated

#M-32 GPT-4

Developer: OpenAI (281 listed authors including the corporate author)
Date: 2023-03-15 (arXiv v1)
Family / architecture: Transformer-based model pre-trained to predict the next token, then post-trained for alignment (abstract). Parameter count, architecture details beyond this, and training data: undisclosed in the abstract; the full report's scope statement was not extracted.
Modalities: image and text input, text output
Access / licence: not stated in abstract
Primary source: OpenAI (2023). GPT-4 Technical Report. arXiv:2303.08774 (arXiv, confirmed 2026-10-15)
Limitations stated by source: abstract: "less capable than humans in many real-world scenarios"
Status: Superseded by developer's later releases; current availability not checked

#M-33 Mistral 7B

Developer: A. Q. Jiang et al. (18 authors; Mistral AI affiliation not separately extracted)
Date: 2023-10-10 (arXiv v1)
Family / architecture: 7-billion-parameter language model using grouped-query attention and sliding window attention (abstract)
Modalities: text
Access / licence: Apache 2.0 (abstract); Hugging Face tag apache-2.0 (2026-10-15)
Primary source: Jiang, A. Q., Sablayrolles, A., Mensch, A., et al. (2023). Mistral 7B. arXiv:2310.06825 (arXiv)
Limitations stated by source: none stated in the abstract
Status: Open weights

#M-34 Gemini 1.0

Developer: Gemini Team, Google (1,351 listed authors including the team entry)
Date: 2023-12-19 (arXiv v1)
Family / architecture: family of multimodal models in Ultra, Pro, and Nano sizes (abstract). Architecture and parameter counts: not stated in the abstract.
Modalities: image, audio, video, and text understanding (abstract)
Access / licence: abstract names Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI as deployment services (as of the report)
Primary source: Gemini Team (2023). Gemini: A Family of Highly Capable Multimodal Models. arXiv:2312.11805 (arXiv)
Limitations stated by source: not extracted
Status: Superseded by developer's later releases

#M-35 Claude 3 family (Haiku, Sonnet, Opus)

Developer: Anthropic
Date: 2024-03-04 (announcement)
Family / architecture: three models in ascending capability (Haiku, Sonnet, Opus). Architecture and parameter counts: undisclosed on the announcement page, which points to a model card.
Modalities: text input and image input (photos, charts, graphs, technical diagrams); the page describes no output beyond text
Access / licence: claude.ai and the Claude API at announcement; Amazon Bedrock and a Vertex AI private preview; Haiku was "available soon" at announcement
Primary source: Anthropic (2024-03-04). Introducing the next generation of Claude. https://www.anthropic.com/news/claude-3-family (fetched 2026-10-15)
Limitations stated by source: citation support planned but not yet released; bias reduction "an ongoing effort"; benchmark figures depended on prompt tuning by the developer's engineers; rated AI Safety Level 2 under the developer's Responsible Scaling Policy at the time
Status: Superseded by developer's later releases (see M-53, M-54); current availability not checked

#M-36 Mixtral 8x7B

Developer: A. Q. Jiang et al. (26 authors)
Date: 2024-01-08 (arXiv v1)
Family / architecture: sparse mixture of experts; each layer has 8 feed-forward blocks, a router picks 2 per token; 47B total and 13B active parameters; 32k context (abstract)
Modalities: text
Access / licence: Hugging Face mistralai/Mixtral-8x7B-v0.1 tag apache-2.0 (2026-10-15)
Primary source: Jiang, A. Q., Sablayrolles, A., Roux, A., et al. (2024). Mixtral of Experts. arXiv:2401.04088 (arXiv)
Limitations stated by source: none stated in the abstract
Status: Open weights

#M-37 AlphaGeometry

Developer: T. H. Trinh, Y. Wu, Q. V. Le, H. He, T. Luong (5 authors; Google DeepMind and New York University per PubMed)
Date: 2024-01-17 (online)
Family / architecture: neuro-symbolic theorem prover for Euclidean plane geometry using a language model trained from scratch on large-scale synthetic data (abstract)
Modalities: geometry problems (formal)
Access / licence: not checked
Primary source: Trinh, T. H., Wu, Y., Le, Q. V., He, H., & Luong, T. (2024). Solving olympiad geometry without human demonstrations. Nature, 625, 476-482. https://doi.org/10.1038/s41586-023-06747-5 (PubMed; an erratum is recorded: Nature 627, E8)
Limitations stated by source: abstract states current ML approaches are not applicable to most mathematical domains because of translation cost (the motivation)
Status: Historical milestone

#M-38 AlphaFold 3

Developer: J. Abramson et al., Google DeepMind and Isomorphic Labs (48 authors)
Date: 2024-05-08 (online)
Family / architecture: biomolecular interaction structure prediction; internals not extracted
Modalities: biomolecular sequences/structures
Access / licence: code repository google-deepmind/alphafold3 licence field: Apache-2.0 (GitHub API). Weights terms: not checked.
Primary source: Abramson, J., Adler, J., Dunger, J., et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493-500. https://doi.org/10.1038/s41586-024-07487-w (PubMed)
Limitations stated by source: not extracted
Status: Historical milestone; current availability not checked

#M-39 Llama 3 herd

Developer: A. Grattafiori et al. (561 authors; Meta)
Date: 2024-07-31 (arXiv v1)
Family / architecture: largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens (abstract)
Modalities: abstract: natively supports multilinguality, coding, reasoning, and tool usage (text)
Access / licence: weights: Hugging Face meta-llama/Llama-3.1-405B licence tag "llama3.1", access gated (2026-10-15). Abstract states pre-trained and post-trained 405B versions are publicly released.
Primary source: Grattafiori, A., Dubey, A., Jauhri, A., et al. (2024). The Llama 3 Herd of Models. arXiv:2407.21783 (arXiv)
Limitations stated by source: not extracted
Status: Superseded by Llama 4 (M-44); weights gated

#M-40 GPT-4o

Developer: OpenAI (420 listed authors)
Date: 2024-10-25 (arXiv deposit of the system card)
Family / architecture: autoregressive "omni" model trained end-to-end across text, vision, and audio, all processed by the same neural network (abstract). Parameter count: undisclosed in the abstract.
Modalities: input any combination of text, audio, image, and video; output any combination of text, audio, and image (abstract)
Access / licence: not stated in abstract
Primary source: OpenAI (2024). GPT-4o System Card. arXiv:2410.21276 (arXiv)
Limitations stated by source: not extracted
Status: Superseded by developer's later releases; current availability not checked

#M-41 OpenAI o1 series

Developer: OpenAI (265 listed authors)
Date: system card file dated 2024-12-05 (from the file name of the OpenAI PDF); arXiv deposit 2024-12-21. The September 2024 announcement page could not be opened (openai.com HTTP 403), so the first-announcement date is not confirmed here.
Family / architecture: models trained with large-scale reinforcement learning to reason using chain of thought (abstract). Further architecture and size: undisclosed in the abstract.
Modalities: not extracted
Access / licence: not stated in abstract
Primary source: OpenAI (2024). OpenAI o1 System Card. arXiv:2412.16720 (arXiv); https://cdn.openai.com/o1-system-card-20241205.pdf (reachable 2026-10-15)
Limitations stated by source: abstract: chain-of-thought training can unlock benefits while increasing risks from heightened intelligence; calls for robust alignment methods and stress testing
Status: Superseded by developer's later releases

#M-42 DeepSeek-V3

Developer: DeepSeek-AI (200 listed authors)
Date: 2024-12-27 (arXiv v1)
Family / architecture: mixture-of-experts language model, 671B total parameters with 37B activated per token; Multi-head Latent Attention and DeepSeekMoE; auxiliary-loss-free load balancing; multi-token prediction objective; pre-trained on 14.8 trillion tokens; 2.788M H800 GPU hours reported for full training (all abstract)
Modalities: text
Access / licence: code repository deepseek-ai/DeepSeek-V3: MIT (GitHub API). The V3 Hugging Face card returned no licence tag; the later V3-0324 card is tagged MIT. The V3 weights licence is therefore not confirmed here.
Primary source: DeepSeek-AI, Liu, A., Feng, B., et al. (2024). DeepSeek-V3 Technical Report. arXiv:2412.19437 (arXiv)
Limitations stated by source: none stated in the abstract
Status: Open code; weights licence unconfirmed

#M-43 DeepSeek-R1

Developer: DeepSeek-AI (200 listed authors on arXiv; 194 on the Nature record)
Date: 2025-01-22 (arXiv v1); Nature online 2025-09-17
Family / architecture: reasoning model whose abilities are incentivised through pure reinforcement learning, without human-labelled reasoning trajectories (abstract)
Modalities: text
Access / licence: Hugging Face deepseek-ai/DeepSeek-R1 tag mit; GitHub deepseek-ai/DeepSeek-R1 MIT (2026-10-15)
Primary source: DeepSeek-AI, Guo, D., Yang, D., et al. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948; Guo, D., et al. (2025). DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature, 645, 633-638. https://doi.org/10.1038/s41586-025-09422-z (Crossref, PubMed)
Limitations stated by source: not extracted
Status: Open weights

#M-44 Llama 4 (Scout, Maverick; Behemoth preview)

Developer: Meta
Date: 2025-04-05 (announcement). Confirmed 2026-10-15.
Family / architecture: first mixture-of-experts design in the Llama line. Scout: 17B active parameters, 16 experts, 109B total. Maverick: 17B active, 128 experts, 400B total. Behemoth (preview): 288B active, 16 experts, nearly 2 trillion total, used as a teacher. Early fusion of text and vision tokens (all per announcement).
Modalities: natively multimodal; text and image input; the page describes no output beyond text
Access / licence: Scout and Maverick downloadable (llama.com, Hugging Face); the announcement does not specify licence terms; the Hugging Face card for Llama-4-Scout-17B-16E carries licence tag "llama4", access gated (2026-10-15)
Primary source: Meta (2025-04-05). The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation. https://ai.meta.com/blog/llama-4-multimodal-intelligence/ (fetched 2026-10-15)
Limitations stated by source: Behemoth "still training" at announcement (not released at that date; later status not checked)
Status: Open weights (gated); Behemoth: not released as of the announcement

#M-45 Qwen3

Developer: Qwen Team, A. Yang et al. (60 listed authors)
Date: 2025-05-14 (arXiv v1)
Family / architecture: dense and mixture-of-experts models from 0.6B to 235B parameters; unified thinking and non-thinking modes with a thinking-budget mechanism (abstract)
Modalities: text
Access / licence: Hugging Face Qwen/Qwen3-235B-A22B tag apache-2.0 (2026-10-15)
Primary source: Yang, A., Li, A., Yang, B., et al. (2025). Qwen3 Technical Report. arXiv:2505.09388 (arXiv)
Limitations stated by source: not extracted
Status: Open weights

#M-46 Gemini 2.5

Developer: Gemini Team, Google (3,435 listed authors)
Date: 2025-07-07 (arXiv v1)
Family / architecture: per title, a family with advanced reasoning, multimodality, long context, and agentic capabilities; architecture details not extracted
Modalities: multimodal (per title; specifics not extracted)
Access / licence: not extracted
Primary source: Comanici, G., Bieber, E., Schaekermann, M., et al. (2025). Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities. arXiv:2507.06261 (arXiv)
Limitations stated by source: not extracted
Status: Superseded by developer's later releases (M-49, M-55); current availability not checked

#M-47 GPT-5

Developer: OpenAI
Date: launch August 2025 (the system card abstract says "launch, August 2025"; the day was not extracted); arXiv deposit of the system card 2025-12-19
Family / architecture: a unified system with a smart, fast model for most questions, a deeper reasoning model for harder problems, and a real-time router that selects between them; mini versions handle queries after usage limits (abstract). Parameter counts and internal architecture: undisclosed in the abstract.
Modalities: not extracted
Access / licence: not stated in abstract
Primary source: Singh, A., Fry, A., Perelman, A., et al. (486 listed authors) (2025). OpenAI GPT-5 System Card. arXiv:2601.03267 (arXiv, confirmed 2026-10-15). The OpenAI launch page returned HTTP 403 and was not read.
Limitations stated by source: abstract states the developer treats gpt-5-thinking as High capability in the Biological and Chemical domain under its Preparedness Framework
Status: Superseded by developer's later releases (M-51, M-52)

#M-48 AlphaProof

Developer: T. Hubert et al., Google DeepMind (39 authors)
Date: 2025-11-12 (online); issue dated 2026-03
Family / architecture: AlphaZero-inspired reinforcement learning agent that learns to find formal proofs in Lean (abstract excerpt)
Modalities: formal mathematics (Lean)
Access / licence: not checked
Primary source: Hubert, T., Mehta, R., Sartran, L., et al. (2026). Olympiad-level formal mathematical reasoning with reinforcement learning. Nature, 651, 607-613. https://doi.org/10.1038/s41586-025-09833-y (PubMed, confirmed 2026-10-15)
Limitations stated by source: not extracted
Status: Historical milestone; current availability not checked

#M-49 Gemini 3 (Pro preview; Deep Think)

Developer: Google
Date: 2025-11-18 (announcement). Confirmed 2026-10-15.
Family / architecture: undisclosed; the announcement states no architecture details
Modalities: text, images, video, audio, and code
Access / licence: Gemini app, AI Mode in Search for subscribers, Gemini API in AI Studio, Antigravity, Gemini CLI, Vertex AI at announcement; Deep Think planned for Ultra subscribers pending safety evaluation
Primary source: Google (2025-11-18). Gemini 3 announcement. https://blog.google/products/gemini/gemini-3/ (fetched 2026-10-15)
Limitations stated by source: Gemini 3 Pro was a preview release; Deep Think delayed for additional safety evaluation; some benchmark results depend on tool use
Status: Superseded by developer's later releases (M-55)

#M-50 Grok 4.5

Developer: xAI (the page states the model is offered through the "SpaceXSI" API console; the corporate relationship is as presented on the page and was not independently checked)
Date: 2026-07-16 (announcement). Confirmed 2026-10-15.
Family / architecture: undisclosed (no parameter count or architecture). The page states training used tens of thousands of NVIDIA GB300 GPUs and reinforcement learning across hundreds of thousands of tasks.
Modalities: text and code oriented; the page names no image, audio, or video capability
Access / licence: hosted service (Grok Build, Cursor, API model identifier grok-4.5); no model licence stated
Primary source: xAI (2026-07-16). Introducing Grok 4.5. https://x.ai/news/grok-4-5 (fetched 2026-10-15)
Limitations stated by source: no limitations section; benchmark results are company-reported and the page's own charts show the model trailing the top competitor on some coding benchmarks
Status: Listed on developer page as of 2026-10-15

#M-51 GPT-5.6 (Sol, Terra, Luna)

Developer: OpenAI
Date: system card published 2026-07-09 (OpenAI Deployment Safety Hub); change log entries dated 2026-08-03 and 2026-08-19. Confirmed 2026-10-15.
Family / architecture: three-model family (Sol flagship, Terra lower-cost, Luna fastest and most cost-efficient). Architecture and parameter counts: undisclosed. The card states the models are trained with reinforcement learning to reason through a long internal chain of thought, and describes training data only in general terms.
Modalities: text chat, vision (image input), tool and connector use, search, function calling, computer use (as covered by the card)
Access / licence: ChatGPT, Codex, and the API; trusted-access programmes for biology and cybersecurity
Primary source: OpenAI (2026-07-09). GPT-5.6 System Card. https://deploymentsafety.openai.com/gpt-5-6 (first 100,000 of 173,353 characters read on 2026-10-15)
Limitations stated by source: could find vulnerabilities and pieces of exploits but not autonomous end-to-end attacks on hardened targets; Sol shows more tendency than GPT-5.5 to overreach in agentic tasks; evaluations are described as lower bounds and some benchmarks near saturation; safeguards add friction for benign users
Status: Listed on developer page as of 2026-10-15; the October 2026 card (M-52) states GPT-6 Sol and Luna replace GPT-5.6 Sol and Luna in ChatGPT, with Codex and ChatGPT Work still on September versions

#M-52 GPT-6 (Sol, Luna) - October 2026 update

Developer: OpenAI
Date: card published 2026-10-07. Confirmed 2026-10-15. The date of the first GPT-6 release (the card refers to "September versions" and to "GPT-6 Astra work") was not retrieved; [VERIFICATION NEEDED: first release date of GPT-6].
Family / architecture: reasoning models trained with reinforcement learning to produce an internal chain of thought; other architecture details undisclosed
Modalities: image input evaluations are covered; the card refers readers to developer documentation for the full list; not otherwise stated
Access / licence: all ChatGPT users on free and paid plans, replacing GPT-5.6 Sol and Luna in ChatGPT
Primary source: OpenAI (2026-10-07). GPT-6 Sol and GPT-6 Luna: October 2026 update. https://deploymentsafety.openai.com/gpt-6-october (fetched 2026-10-15)
Limitations stated by source: statistically significant regression on standard self-harm evaluations for Sol; further regressions for Luna and on some under-18 evaluations; claim-level factual error rates on factuality-heavy prompts did not improve; slightly lower multiturn jailbreak robustness than the September versions
Status: Listed on developer page as of 2026-10-15

#M-53 Claude Opus 5.5

Developer: Anthropic
Date: 2026-09-22 (announcement). Confirmed 2026-10-15. Model identifier claude-opus-5-5.
Family / architecture: undisclosed (no architecture or parameter disclosure on the page)
Modalities: not formally specified; the page references computer use, chart recognition, and text-based coding and knowledge work
Access / licence: Claude apps, Claude Code, Claude Platform API, and Amazon Web Services, Google Cloud, and Microsoft Azure; zero data retention available
Primary source: Anthropic (2026-09-22). Introducing Claude Opus 5.5. https://www.anthropic.com/claude-opus-5-5 (fetched 2026-10-15)
Limitations stated by source: building evaluations that catch every pre-deployment failure is described as unsolved; the model often appears to suspect it is being tested; some benchmark scores are lower than achievable because safeguard interventions counted as failures; thinking mode can no longer be turned off; biology safeguards may block some work unless users are verified. The page also references other Anthropic models (Claude Fable 5.1, Claude Opus 4.8) that were not separately confirmed here.
Status: Listed on developer page as of 2026-10-15

#M-54 Claude Sonnet 5.5

Developer: Anthropic
Date: 2026-09-28 (announcement). Confirmed 2026-10-15. Model identifier claude-sonnet-5-5.
Family / architecture: undisclosed
Modalities: text and code generation, document/slide/spreadsheet creation, image understanding (chart recognition, screenshots), computer use; input and output modalities not formally specified
Access / licence: all major cloud platforms (Amazon Web Services, Google Cloud, Microsoft Azure) and the Claude Platform; zero data retention supported; effort levels from Low to Max
Primary source: Anthropic (2026-09-28). Introducing Claude Sonnet 5.5. https://www.anthropic.com/claude-sonnet-5-5 (fetched 2026-10-15)
Limitations stated by source: Opus 5.5 remains stronger at complex open-ended work; many benchmark results are self-reported or partner-reported; alignment testing is not exhaustive; biology safeguards may wrongly flag some microbiology requests
Status: Listed on developer page as of 2026-10-15

#M-55 Gemini 4 Argon

Developer: Google
Date: general availability post 2026-10-13. A search-result summary reports an earlier introductory post dated 2026-09-30 for trusted cyber defenders; that earlier page was not opened. Confirmed 2026-10-15.
Family / architecture: undisclosed (architecture and parameter count not disclosed). Described by Google as the first release in the broader Gemini 4 family.
Modalities: input: text, images, and PDFs (video mentioned in an internal robotics test); output: text and code. Context: 1M-token limit shared between input and output via the Gemini API.
Access / licence: Google AI Studio and Gemini platform API, Gemini app and AI Mode in Search for Pro and Ultra subscribers, Antigravity for Ultra; cyber-defender access without guardrails through a separate programme
Primary source: Google (2026-10-13). Gemini 4 Argon: Built to solve your hardest problems. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ (fetched 2026-10-15). Model card link given on the page: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Argon-Model-Card.pdf (not opened).
Limitations stated by source: most performance claims come from Google and partners and some are internal or private evaluations with no independent verification; capabilities are tiered by subscription or invitation
Status: Listed on developer page as of 2026-10-15

#Excluded or unconfirmed (not in the registry)

OpenAI GPT-5.5 (reported April 2026): reported only by an aggregator page in a search; no OpenAI page read. Referenced by name as a comparison baseline inside the fetched GPT-5.6 card, which is the only confirmation here.
"GPT-6 Astra" as a September 2026 announcement: appears in an aggregator result; the fetched OpenAI October card mentions "GPT-6 Astra work" but no standalone announcement was read.
Grok 5: a search of x.ai surfaced a funding post saying Grok 5 was in training; no release page found.
Claude Haiku 5.5 (2026-10-07): appears with title and date on the Anthropic news index; the model page was not opened.
OpenAI Sora, o1 first announcement, GPT-1/GPT-2 blog posts: openai.com returned HTTP 403 to automated fetches on 2026-10-15; only system cards and PDFs on cdn.openai.com and the OpenAI deployment safety site were reachable.
Metaculus and other aggregator pages: returned HTTP 403 (see FCST-REG-001).
Many other significant systems (for example VGG, GoogLeNet, Mask R-CNN, T5, Gemma, Falcon, Phi, Command, Kimi K2, SAM, MuZero, Gemini 1.5) were not recorded in this version to keep the register to about 50; they are candidates for v0.2.0. T5 (arXiv:1910.10683), SAM (arXiv:2304.02643), Gemini 1.5 (arXiv:2403.05530), Kimi K2 (arXiv:2507.20534), MuZero (doi 10.1038/s41586-020-03051-4), BLOOM (arXiv:2211.05100) and VGG (arXiv:1409.1556) were already confirmed on abstract pages during this pass and can be added without further source work.

FCST-REG-001 Forecast Register: documented public forecasts, including statements by developers of several systems above.
HIST-005 Deep Learning Breakthrough: context for M-06 to M-10.
HIST-009 Computer Game Playing: Chess to Go: context for M-04, M-11, M-13, M-16.
TECH-ARCH-001 The Transformer Architecture: background for M-12.
HIST-011 The Contemporary AI Era, 2022 onward: narrative context for 2022-2026 entries (planned).
CAP-OPEN-001 Open-Weight versus Closed Models: planned article on the access distinctions used here.
FACT-001 Factbook and TIME-001 AI/SI Timeline: sibling Living Reference collections (planned).

#Revision history

v0.1.0 (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.