Models (198)

OpenAI (26)

  • GPT-1 · OpenAI's June 2018 Generative Pre-trained Transformer, the 117-million-parameter model that established the pretrain-then-finetune recipe behind the GPT series.
  • GPT-2 · OpenAI's 1.5-billion-parameter 2019 model; demonstrated multitask behavior from pure language modeling and started the staged-release debate.
  • GPT-3 · OpenAI's 175-billion-parameter language model of 2020, which demonstrated few-shot in-context learning at scale.
  • CLIP · OpenAI's January 2021 contrastive vision-language model that learns image classification from natural-language supervision and became a standard component in text-to-image systems.
  • DALL-E · OpenAI's January 2021 text-to-image model, a 12-billion-parameter autoregressive Transformer that generated images from natural-language prompts.
  • Codex (2021 model) · OpenAI's 2021 code-generation model, a GPT-3 descendant fine-tuned on public code; the original engine of GitHub Copilot and source of the HumanEval benchmark.
  • InstructGPT · OpenAI's 2022 RLHF-aligned GPT-3 variants; demonstrated that alignment training beats scale on human preference and set the template for ChatGPT.
  • DALL-E 2 · OpenAI's April 2022 text-to-image diffusion model, which generated 1024x1024 images from CLIP embeddings and popularized AI image generation.
  • Whisper · OpenAI's open-source speech recognition model of September 2022, trained on 680,000 hours of weakly supervised multilingual audio.
  • GPT-3.5 · OpenAI's 2022 model series bridging GPT-3 and GPT-4; the original engine of ChatGPT.
  • GPT-4 · OpenAI's fourth-generation GPT model, released March 14, 2023; multimodal input and undisclosed architecture.
  • GPT-4 Turbo · OpenAI's November 2023 update to GPT-4 with a 128,000-token context window, fresher training data, and substantially lower API pricing.
  • DALL-E 3 · OpenAI's September 2023 text-to-image model, built to follow detailed prompts closely and integrated natively into ChatGPT.
  • Sora · OpenAI's text-to-video generation model, previewed in February 2024 and built on a diffusion transformer over spacetime patches.
  • o3 · OpenAI's frontier reasoning model, previewed in December 2024 and released in April 2025, noted for developer-reported gains on ARC-AGI, math, and coding benchmarks.
  • GPT-4o · OpenAI's May 2024 'omni' model: natively multimodal across text, vision, and audio, with real-time voice; became ChatGPT's long-running default.
  • GPT-4o mini · OpenAI's July 2024 small multimodal model, a low-cost successor to GPT-3.5 Turbo in ChatGPT and the API.
  • o1 · OpenAI's first reasoning model (September 2024): trained with RL to think in long chains before answering, opening the test-time compute era.
  • GPT-4.5 · OpenAI's February 2025 research-preview chat model, positioned as its largest scaling-focused (non-reasoning) model and retired from the API in mid-2025.
  • o4-mini · OpenAI's April 2025 small reasoning model, released alongside o3 with tool use and image-based reasoning at a lower price point.
  • gpt-oss · OpenAI's August 2025 open-weight models (120B and 20B MoE, Apache 2.0), its first open weights since GPT-2.
  • GPT-5 · OpenAI's August 2025 flagship: a unified system routing between fast and reasoning models, ending the GPT-4-era model picker.
  • Sora 2 · OpenAI's September 2025 video-and-audio generation model, launched alongside a consumer Sora app with likeness-based 'cameos'.
  • GPT-5.1 · OpenAI's November 2025 update to GPT-5, pairing a warmer default conversational style with adaptive reasoning in Instant and Thinking variants.
  • GPT-5.2 · OpenAI's December 2025 update to the GPT-5 series, shipped in Instant, Thinking, and Pro variants amid intensified competition with Google's Gemini 3.
  • GPT-5.6 · OpenAI's July 2026 frontier model series (Sol, Terra, Luna), first released as a restricted preview to US-government-vetted organizations in June 2026.

Google DeepMind (23)

  • AlphaGo · DeepMind's Go system: deep policy and value networks guiding tree search, the first program to beat a Go professional (2015) and 4-1 winner over Lee Sedol in March 2016.
  • AlphaZero · DeepMind's general self-play algorithm (2017) that mastered chess, shogi, and Go from the rules alone, defeating Stockfish, Elmo, and AlphaGo Zero without human data.
  • AlphaFold · DeepMind's protein structure prediction line: CASP13 winner (2018), CASP14's effective solution of single-chain folding via AlphaFold 2 (2020), and AlphaFold 3 (2024) for biomolecular complexes.
  • Meena · Google's January 2020 open-domain neural chatbot, a 2.6-billion-parameter conversational model that preceded LaMDA.
  • LaMDA · Google's 2021 dialogue-specialized language model family, a predecessor of Bard and the Gemini line, and the subject of a widely covered 2022 sentience controversy.
  • Gopher · DeepMind's December 2021 280-billion-parameter dense language model, the baseline that motivated the Chinchilla compute-optimal scaling analysis.
  • Chinchilla · DeepMind's 2022 compute-optimal model: 70B parameters trained on 1.4T tokens, proving the GPT-3 generation was undertrained and resetting industry training budgets.
  • Flamingo · DeepMind's April 2022 visual language model that pioneered few-shot multimodal prompting by bridging frozen vision and language models.
  • PaLM · Google's 540-billion-parameter Pathways Language Model of April 2022, a landmark in dense scaling and few-shot reasoning.
  • Imagen · Google's text-to-image diffusion model line, introduced in May 2022 and continued through Imagen 2, 3, and 4.
  • Flan-T5 · Google's October 2022 family of instruction-finetuned T5 models, released openly in five sizes and a standard baseline for instruction-tuning research.
  • PaLM 2 · Google's May 2023 large language model, successor to PaLM, that powered Bard and Workspace features before being replaced by Gemini.
  • SynthID · Google DeepMind's watermarking system for labeling AI-generated images, audio, video, and text, first announced in August 2023.
  • Gemini 1.0 · Google DeepMind's December 2023 natively multimodal model family (Ultra, Pro, Nano), the first flagship release of the Gemini line.
  • Genie · Google DeepMind's series of generative interactive world models (Genie, Genie 2, Genie 3) that produce playable environments from images or text.
  • Gemini 1.5 · Google DeepMind's February 2024 generation; made million-token context windows a product reality and introduced the Flash tier.
  • Gemma · Google DeepMind's open-weight line drawing on Gemini research: Gemma 1 (2024) through the multimodal Gemma 3 (2025).
  • PaliGemma · Google's open vision-language model of May 2024, pairing a SigLIP vision encoder with a Gemma language decoder and designed for fine-tuning on downstream tasks.
  • Veo · Google DeepMind's flagship text-to-video generation model series, first announced in May 2024 and extended with native audio in Veo 3.
  • Gemini 2.0 · Google DeepMind's December 2024 'agentic era' generation: Flash-first release, native tool use, and the Flash Thinking reasoning experiments.
  • Gemini 2.5 · Google DeepMind's 2025 thinking-by-default generation; 2.5 Pro led leaderboards through spring 2025.
  • Gemini 3 · Google DeepMind's November 2025 generation; launch-led major benchmarks and deepened the Deep Think reasoning tier.
  • Gemini 3.5 · Google DeepMind's 2026 Gemini generation: Flash shipped at I/O in May 2026 as the default Gemini model, while the Pro flagship remained unreleased as of July 2026.

Anthropic (17)

  • Claude 1 · Anthropic's first publicly released chat assistant (March 2023), notable for its Constitutional AI alignment approach and later 100K-token context window.
  • Claude Instant · Anthropic's lighter, faster, lower-cost assistant line launched alongside the original Claude in March 2023 and retired after the Claude 3 generation introduced Haiku.
  • Claude 2 · Anthropic's July 2023 model; its 100K-token context window was the era's longest, and its cautious refusals defined early Claude's reputation.
  • Claude 3 · Anthropic's March 2024 generation (Haiku, Sonnet, Opus); Opus was the first model to consistently challenge GPT-4-class results.
  • Claude 3.5 Sonnet · Anthropic's mid-2024 workhorse; beat Claude 3 Opus at twice the speed, and its October refresh introduced computer use.
  • Claude 3.5 Haiku · Anthropic's fast, low-cost model of the Claude 3.5 generation, announced in October 2024 as the successor to Claude 3 Haiku.
  • Claude 3.7 Sonnet · Anthropic's February 2025 hybrid reasoning model: one network serving instant and extended-thinking modes, launched alongside Claude Code.
  • Claude 4 · Anthropic's May 2025 generation (Opus 4, Sonnet 4), built for agentic coding; Opus 4 was the first model deployed under ASL-3 safeguards.
  • Claude Opus 4.1 · Anthropic's August 2025 incremental upgrade to Claude Opus 4, with developer-reported gains in agentic coding and long-horizon reasoning.
  • Claude Sonnet 4.5 · Anthropic's September 2025 mid-tier hybrid reasoning model, marketed at launch as the company's strongest model for coding and computer use.
  • Claude Haiku 4.5 · Anthropic's October 2025 small model, the first Haiku with extended thinking, offering developer-reported near-Sonnet-4 coding performance at a fraction of the cost.
  • Claude Opus 4.5 · Anthropic's November 2025 flagship: developer-reported coding leadership with a major price cut, closing the Claude 4.x series.
  • Claude Opus 4.6 · Anthropic's February 2026 flagship, adding a beta 1M-token context window, adaptive thinking, and effort controls to the Opus line.
  • Claude Opus 4.8 · Anthropic's May 2026 Opus flagship, adding dynamic workflows, a discounted fast mode, and developer-reported gains in coding honesty; the last Opus release before Claude Fable 5.
  • Claude Fable 5 · Anthropic's June 2026 frontier model, first of the Mythos-class tier above Opus; briefly suspended under a US export-control directive.
  • Claude Sonnet 5 · Anthropic's June 2026 mid-tier model, marketed as its most agentic Sonnet, released as the default model for Free and Pro plans.
  • Claude Opus 5 · Anthropic's July 2026 Opus release, positioned as approaching Claude Fable 5's capability at half its price; the first Opus-tier model of the Claude 5 generation.

Meta AI (12)

  • RoBERTa · Facebook AI's July 2019 replication study of BERT that showed the original model was significantly undertrained, setting new benchmark records with an optimized pretraining recipe.
  • OPT · Meta AI's May 2022 suite of open decoder-only language models, from 125M to 175B parameters, released with weights, code, and a candid training logbook.
  • LLaMA · Meta's February 2023 research models (7B-65B); their prompt leak seeded the open-source LLM ecosystem.
  • Segment Anything · Meta AI's April 2023 promptable image segmentation model (SAM), released with the 1.1-billion-mask SA-1B dataset.
  • Llama 2 · Meta's July 2023 openly licensed models (7B-70B) with RLHF chat variants; made open weights an official corporate strategy.
  • Code Llama · Meta AI's August 2023 family of open-weight code-specialized language models derived from Llama 2, released in 7B to 70B sizes.
  • Llama 3 · Meta's 2024 generation: 8B/70B in April, the 405B frontier-scale release in July; the herd paper documented open training at frontier scale.
  • Llama 3.1 · Meta AI's July 2024 open-weight model family (8B, 70B, 405B), whose 405B variant was widely described as the first open-weight model competitive with frontier closed systems.
  • Llama 3.2 · Meta AI's September 2024 Llama release adding small on-device text models (1B, 3B) and the family's first vision-capable models (11B, 90B).
  • Movie Gen · Meta AI's October 2024 suite of media foundation models for text-to-video and synchronized audio generation, shown as a research preview without a public release.
  • Llama 3.3 · Meta AI's December 2024 open-weight 70B instruct model, developer-reported to approach Llama 3.1 405B quality at a fraction of the serving cost.
  • Llama 4 · Meta's April 2025 MoE generation (Scout, Maverick; Behemoth unreleased); a mixed reception that preceded Meta's Superintelligence Labs reorganization.

Mistral AI (10)

  • Mixtral 8x7B · Mistral AI's December 2023 sparse MoE, the first widely deployed open mixture-of-experts model; dropped as a magnet link before any announcement.
  • Mistral 7B · Mistral AI's September 2023 debut model: a 7.3-billion-parameter open-weight language model that outperformed larger Llama 2 variants.
  • Mistral Small · Mistral AI's mid-size model line, launched February 2024 and later released open-weight, positioned for low-latency deployment below Mistral Large.
  • Mistral Large · Mistral AI's flagship proprietary language model line, launched February 2024 and updated in July 2024 as the 123-billion-parameter Mistral Large 2.
  • Mixtral 8x22B · Mistral AI's April 2024 sparse mixture-of-experts model with about 141B total and 39B active parameters, released under Apache 2.0.
  • Codestral · Mistral AI's May 2024 code-generation model, a 22-billion-parameter weight-available system trained on more than 80 programming languages.
  • Pixtral · Mistral AI's first natively multimodal model line, opened in September 2024 with the Apache-2.0 Pixtral 12B vision-language model.
  • Devstral · Mistral AI's open-weight agentic coding model line, launched May 2025 with All Hands AI and tuned for software-engineering agent workflows.
  • Mistral Medium 3 · Mistral AI's May 2025 proprietary enterprise model, marketed as near-frontier performance at a substantially lower price.
  • Magistral · Mistral AI's first reasoning model line, released in June 2025 as the open-weight Magistral Small (24B) and the proprietary Magistral Medium.

Qwen team (10)

  • Qwen 1 · Alibaba Cloud's first generation of Qwen large language models (2023), released as open-weight checkpoints from 1.8B to 72B parameters.
  • QwQ-32B · Alibaba Qwen team's open-weight 32B reasoning model, previewed in November 2024 and released in March 2025 with developer-reported parity to much larger reasoners.
  • Qwen2 · Alibaba's June 2024 open-weight model series spanning 0.5B to 72B parameters, including one mixture-of-experts variant.
  • Qwen2.5 · Alibaba's September 2024 open-weight model series (0.5B to 72B), trained on a developer-reported 18 trillion tokens and widely used as a base for fine-tunes and reasoning distillations.
  • Qwen2.5-Max · Alibaba's January 2025 proprietary large-scale mixture-of-experts flagship, positioned against DeepSeek-V3 and Western frontier models.
  • Qwen2.5-VL · Alibaba's January 2025 open-weight vision-language model series (3B to 72B) with document parsing, object grounding, long-video understanding, and GUI-agent capabilities.
  • Qwen3 · Alibaba's April 2025 open-weight model family introducing hybrid thinking and non-thinking modes across dense and mixture-of-experts variants.
  • Qwen3-Coder · Alibaba's July 2025 open-weight agentic coding model, a 480B-parameter mixture-of-experts specialist derived from the Qwen3 family.
  • Qwen3-Max · Alibaba's September 2025 proprietary flagship of the Qwen3 generation, developer-reported to exceed one trillion parameters.
  • Qwen3-Coder-Next · Alibaba Qwen team's February 2026 open-weight coding model, an ultra-sparse 80B hybrid-attention MoE tuned for coding agents and local development.

xAI (7)

  • Grok-1 · xAI's first model (November 2023); its March 2024 open-weights drop, a 314B MoE under Apache 2.0, was then the largest open release.
  • Grok-1.5 · xAI's March 2024 successor to Grok-1, adding a 128,000-token context window and developer-reported gains in math and coding.
  • Grok-2 · xAI's August 2024 frontier chatbot model, released in beta on the X platform alongside a smaller Grok-2 mini variant.
  • Grok 3 · xAI's February 2025 flagship model, trained on the Colossus supercluster and released with a dedicated reasoning mode and the DeepSearch agent.
  • Grok 4 · xAI's July 2025 flagship reasoning model, trained with large-scale reinforcement learning and released alongside a multi-agent Grok 4 Heavy variant.
  • Grok 4.1 · xAI's November 2025 update to Grok 4, focused on emotional intelligence, creative writing, and reduced hallucinations, with a developer-reported first place on the LMArena text leaderboard at launch.
  • Grok 4.5 · xAI's July 2026 frontier model, pitched as an efficient coding and agentic workhorse with configurable reasoning effort.

DeepSeek (6)

  • DeepSeek LLM · DeepSeek's first general-purpose open-weight language model series (7B and 67B), released in late 2023 and accompanied by a scaling-laws study.
  • DeepSeek-V2 · DeepSeek's May 2024 open MoE that introduced multi-head latent attention and DeepSeekMoE; started China's LLM price war.
  • DeepSeek-Coder-V2 · DeepSeek's June 2024 open-weight mixture-of-experts code model, developer-reported as competitive with closed frontier models on coding benchmarks.
  • DeepSeek-V3 · DeepSeek's December 2024 open-weight flagship: 671B-parameter MoE reporting frontier quality from a ~$5.6M disclosed final training run.
  • DeepSeek-R1 · DeepSeek's January 2025 open-weight reasoning model; matched o1-class results with a published RL recipe and triggered a global market repricing.
  • DeepSeek-V3.1 · DeepSeek's August 2025 hybrid open-weight model that merged thinking and non-thinking modes into a single checkpoint, succeeding both DeepSeek-V3 and DeepSeek-R1.

Zhipu AI (6)

  • GLM-130B · A 130-billion-parameter open bilingual (English-Chinese) language model released by Tsinghua University and Zhipu AI in August 2022.
  • ChatGLM · Zhipu AI and Tsinghua's 2023 bilingual open-weight chat assistant line, launched with the locally runnable ChatGLM-6B.
  • GLM-4 · Zhipu AI's January 2024 flagship language model generation, spanning a proprietary API flagship and later open-weight GLM-4-9B variants.
  • GLM-4.5 · Zhipu AI's July 2025 open-weight mixture-of-experts model, positioned as a unified agentic, reasoning, and coding foundation model.
  • GLM-4.6 · Zhipu AI's late-2025 open-weight mixture-of-experts model, an agentic-coding-focused successor to GLM-4.5 with a 200K-token context window.
  • GLM-5 · Zhipu AI's February 2026 open-weight frontier model: a 744B-parameter mixture-of-experts LLM released under the MIT license, followed by the GLM-5.2 update in June 2026.

Arcee AI (5)

  • AFM-4.5B · The first from-scratch Arcee Foundation Model: a 4.5B-parameter enterprise-focused open-weight LLM trained on 8 trillion tokens, later extended to a 64K context window.
  • Trinity Mini · Arcee AI's 26B-parameter sparse mixture-of-experts model (3B active) for agents and tool orchestration, released open-weight under Apache 2.0 in December 2025.
  • Trinity Nano · Arcee AI's 6B-parameter sparse mixture-of-experts model (about 1B active per token) with a 128K context window, aimed at edge, embedded, and offline deployments.
  • Trinity Large · Arcee AI's 400B-parameter sparse mixture-of-experts flagship, trained from scratch on 17 trillion tokens and released in January 2026; widely described as the largest open-weight model from a US lab to that date.
  • Trinity Large Thinking · Arcee AI's April 2026 open-weight reasoning model: an Apache 2.0 variant of the 400B-parameter Trinity Large, post-trained with SFT and RL for long-horizon agent and tool-use workloads.

EleutherAI (5)

  • GPT-Neo · EleutherAI's March 2021 family of open-source autoregressive language models (up to 2.7 billion parameters), an early public replication of the GPT-3 design.
  • GPT-J · EleutherAI's June 2021 open-weight 6-billion-parameter autoregressive language model, an early open alternative to GPT-3-class systems.
  • GPT-NeoX-20B · EleutherAI's February 2022 open-source 20-billion-parameter autoregressive language model, at release the largest publicly available dense model with fully open weights.
  • Pythia · EleutherAI's April 2023 suite of 16 open-weight language models (70M-12B parameters) with released training checkpoints, built for research on training dynamics and interpretability.
  • Comma (models) · Pair of 7-billion-parameter open-weight language models released in June 2025 by EleutherAI and Common Pile collaborators, trained solely on the openly licensed Common Pile v0.1 corpus.

Hugging Face (5)

  • DistilBERT · Hugging Face's 2019 distilled version of BERT: roughly 40% smaller and 60% faster while retaining about 97% of BERT's language-understanding performance.
  • BLOOM · BigScience's 176-billion-parameter open-access multilingual language model, released in July 2022 as a large-scale collaborative science project coordinated by Hugging Face.
  • StarCoder · BigCode's open code models (2023-2024), trained on the consent-audited Stack corpus; the open-governance benchmark for code LLMs.
  • IDEFICS · Hugging Face's August 2023 open reproduction of DeepMind's Flamingo: an 80B open-weight visual language model trained only on public data.
  • SmolLM · Hugging Face's family of small open-weight language models (135M to 3B parameters), launched in July 2024 and built on openly documented training corpora.

Ai2 (4)

  • ELMo · 2018 contextualized word embedding model from AI2 and the University of Washington, built on a bidirectional LSTM language model.
  • OLMo · AI2's fully open model line: weights, data, code, and checkpoints all public; the scientific control group of the LLM era.
  • Molmo · Ai2's September 2024 family of open-weight vision-language models trained on the human-annotated PixMo dataset rather than synthetic captions distilled from proprietary systems.
  • Tulu 3 · Ai2's November 2024 family of openly post-trained Llama 3.1 derivatives, released with its full recipe and known for introducing reinforcement learning with verifiable rewards (RLVR).

MiniMax (4)

  • Hailuo · MiniMax's consumer AI brand, covering the Hailuo AI assistant app and the Hailuo video-generation model line (video-01, Hailuo 02).
  • MiniMax-Text-01 · MiniMax's January 2025 open-weight language model, notable for its hybrid lightning-attention architecture and developer-reported 4-million-token context window.
  • MiniMax-M1 · MiniMax's June 2025 open-weight hybrid-attention reasoning model, developer-reported as the first large-scale open-weight model combining lightning attention with mixture-of-experts.
  • MiniMax-M2 · MiniMax's October 2025 open-weight mixture-of-experts model aimed at agentic and coding workflows, released under the MIT license.

ByteDance (3)

  • Doubao · ByteDance's assistant model line; China's largest chatbot user base, priced to force market-wide cuts.
  • Seedream · ByteDance Seed's text-to-image generation model family (Seedream 2.0-4.0), noted for bilingual Chinese-English prompting and integrated image editing.
  • Seed-OSS · ByteDance Seed's August 2025 open-weight 36B dense language model family, notable for a long native context window and a user-controllable thinking budget.

Microsoft AI (3)

  • Turing-NLG · Microsoft's 17-billion-parameter language model of February 2020, briefly the largest published language model and the public debut of the DeepSpeed library.
  • LLaVA · The 2023 open vision-language recipe: a frozen vision encoder bridged to an open LLM with GPT-4-generated instruction data.
  • MAI-1 · Microsoft AI's in-house foundation model line: the mixture-of-experts MAI-1-preview entered public testing on LMArena in August 2025 and began feeding Copilot text features.

Moonshot AI (3)

  • Kimi k1.5 · Moonshot AI's January 2025 multimodal reasoning model, trained with long-context reinforcement learning and released the same week as DeepSeek-R1.
  • Kimi K2 · Moonshot AI's July 2025 trillion-parameter open-weight MoE; put open weights at the agentic frontier and introduced the MuonClip optimizer.
  • Kimi K3 · Moonshot AI's July 2026 flagship: a 2.8-trillion-parameter MoE, the largest open-weight release to date, with the Kimi Delta Attention architecture.

NVIDIA (3)

  • Megatron-LM · NVIDIA's 2019 research project and open-source framework for training multi-billion-parameter Transformer language models with model parallelism.
  • Megatron-Turing NLG · A 530-billion-parameter dense language model trained jointly by Microsoft and NVIDIA, announced in October 2021 as the largest dense Transformer of its time.
  • Nemotron · NVIDIA's family of open-weight language models, best known for the June 2024 Nemotron-4 340B release aimed at synthetic data generation.

Tencent (3)

  • Hunyuan · Tencent's model line: the open Hunyuan-Large MoE (2024), the leading open video model HunyuanVideo, and the Apache-2.0 Hunyuan 3.0 (2026).
  • HunyuanVideo · Tencent's December 2024 open-weight text-to-video generation model, a roughly 13-billion-parameter diffusion transformer released with code and weights.
  • Hunyuan 3.0 · Tencent's July 2026 open-weight flagship: a 295B-parameter mixture-of-experts model with 21B active parameters and a 256K context window, released under Apache 2.0.

AI21 Labs (2)

  • Jurassic (models) · AI21 Labs' family of large autoregressive language models, spanning Jurassic-1 (2021) and Jurassic-2 (2023), later succeeded by Jamba.
  • Jamba · AI21 Labs' 2024 hybrid: the first production-scale model interleaving Transformer attention with Mamba state-space layers.

Baidu (2)

  • ERNIE Bot · Baidu's conversational AI service and flagship LLM line, launched in March 2023 as China's first major ChatGPT-style chatbot.
  • ERNIE 5 · Baidu's fifth-generation ERNIE flagship series: the natively omni-modal ERNIE 5.0 (November 2025) and the efficiency-focused ERNIE 5.1 (May 2026).

Cohere (2)

  • Command (models) · Cohere's enterprise model line; Command R (2024) made retrieval-augmented generation and tool use the product, and Command A (2025) optimized for two-GPU serving.
  • Aya · Cohere For AI's open multilingual model family (2024-2025), built with thousands of volunteer contributors to extend instruction-tuned LLMs beyond English.

Shanghai AI Laboratory (2)

  • InternVL · Shanghai AI Laboratory's open-weight vision-language model family, launched December 2023 and iterated through InternVL 1.5, 2, 2.5, and 3.
  • InternLM · Shanghai AI Laboratory's open-weight language model series, spanning the original 2023 release through InternLM2, InternLM2.5, and InternLM3.

Yandex (2)

  • YaLM-100B · Yandex's June 2022 open 100-billion-parameter GPT-like language model, at release the largest dense model with freely downloadable Apache 2.0 weights.
  • YandexGPT · Family of large language models by Yandex, launched May 2023 inside the Alice assistant and served via the Yandex Cloud API, with openly released 8B checkpoints.

01.AI (1)

  • Yi (model family) · 01.AI's family of bilingual open-weight language models, launched in November 2023 with Yi-6B and Yi-34B and later extended by Yi-1.5 and the proprietary Yi-Large.

Amazon (AGI team) (1)

  • Amazon Nova · Amazon's in-house foundation-model family (December 2024): the Micro/Lite/Pro ladder plus creative models, distributed exclusively through Bedrock.

Amazon (AWS) (1)

  • Amazon Titan · Amazon's first-party foundation-model family (announced April 2023), served exclusively through Amazon Bedrock and later superseded by Amazon Nova.

Ant Group (1)

  • Ling (models) · Ant Group's open-weight mixture-of-experts model family (Bailing line), notable for frontier-scale training partly on Chinese domestic accelerators.

Apple (1)

  • Apple foundation models · Apple's family of proprietary language models (an approximately 3B-parameter on-device model plus a larger server model) powering Apple Intelligence, announced June 2024.

Black Forest Labs (1)

  • FLUX · Black Forest Labs' family of rectified flow transformer text-to-image models, launched in August 2024 with open-weight and proprietary variants.

Cerebras Systems (1)

  • Cerebras-GPT · Cerebras's March 2023 family of seven open compute-optimal GPT models (111M to 13B), trained on The Pile to demonstrate its wafer-scale hardware.

Databricks (1)

  • DBRX · Databricks' March 2024 open-weight mixture-of-experts language model, with 132 billion total and 36 billion active parameters.

Google (1)

  • BERT · Google's 2018 encoder-only Transformer, pretrained with masked language modeling; defined the pretrain-then-finetune era in NLP.

Google Brain and Carnegie Mellon University (1)

  • XLNet · A June 2019 generalized autoregressive pretraining model from Google Brain and Carnegie Mellon University that outperformed BERT on many benchmarks.

Google Research (Brain team) (1)

  • T5 · Google's 2019 Text-to-Text Transfer Transformer, an open-weight encoder-decoder family that cast every NLP task as text-to-text.

Google Research, with Stanford University (1)

  • ELECTRA · A 2020 Google Research and Stanford encoder model pre-trained with replaced token detection, reaching BERT-class quality at a fraction of the compute.

Google Research; Toyota Technological Institute at Chicago (1)

  • ALBERT · A 2019 parameter-efficient variant of BERT from Google Research that used factorized embeddings and cross-layer parameter sharing to cut model size.

IBM (IBM Research) (1)

  • IBM Granite · IBM's family of open-weight enterprise language models, spanning dense, mixture-of-experts, code, and hybrid Mamba/transformer variants.

IBM Research (Gerald Tesauro) (1)

  • TD-Gammon · Gerald Tesauro's IBM backgammon program that taught itself master-level play by temporal-difference self-play, the first landmark success of reinforcement learning on a complex task.

iFlytek (1)

  • iFlytek Spark · iFlytek's family of large language models, launched in May 2023 and notable for being trained on domestic Huawei Ascend compute.

Ilya Sutskever (1)

  • AlexNet · Krizhevsky, Sutskever, and Hinton's 2012 GPU-trained deep convolutional network whose runaway ILSVRC win (15.3 percent top-5 error) launched the deep learning era in computer vision.

Jeremy Howard (fast.ai) and Sebastian Ruder (1)

  • ULMFiT · Howard and Ruder's 2018 three-stage transfer-learning recipe for text, a direct precursor to the pretrain-and-fine-tune paradigm of BERT and GPT.

Johns Hopkins University (Terence Sejnowski and Charles Rosenberg) (1)

  • NETtalk · Sejnowski and Rosenberg's 1986 network that learned to pronounce English text by backpropagation, the parallel distributed processing era's most public demonstration that multi-layer networks could learn a realistic cognitive task.

Kuaishou Technology (1)

  • Kling · Kuaishou's text-to-video generation model, launched in June 2024 as one of the first widely accessible answers to OpenAI's Sora.

Liquid AI (1)

  • LFM (models) · Liquid AI's model line: the proprietary non-Transformer LFM1 generation (2024) and the open-weight LFM2/LFM2.5 edge families built on hybrid convolution-attention blocks.

LMSYS team (UC Berkeley, CMU, Stanford, UC San Diego) (1)

  • Vicuna · March 2023 open chatbot from the LMSYS team, fine-tuned from LLaMA on shared ChatGPT conversations and a catalyst for LLM-as-a-judge evaluation.

Meituan (1)

  • LongCat · Meituan's open MoE line: LongCat-Flash (560B, 2025) and LongCat-2.0 (1.6T, 2026), trained entirely on Chinese accelerators.

MosaicML (1)

  • MPT (models) · MosaicML's 2023 open-weight MPT line (7B and 30B), an early commercially usable alternative to research-only LLaMA licenses.

RWKV community project (Bo Peng), under the Linux Foundation AI (1)

  • RWKV · Community-built linear-attention RNN trained like a Transformer; the open ecosystem's longest-running alternative-architecture project.

Salesforce Research (1)

  • BLIP · Salesforce Research's vision-language pre-training family (2022-2023), whose BLIP-2 Q-Former popularized bridging frozen image encoders to frozen LLMs.

SenseTime (1)

  • SenseNova · SenseTime's foundation model family, launched in April 2023 and iterated through multimodal and reasoning-focused versions.

Stability AI (1)

  • Stable Diffusion · Open-weight latent diffusion text-to-image model family released in August 2022 by Stability AI with CompVis and Runway, spanning v1 through Stable Diffusion 3.5.

Stanford CRFM (Taori, Gulrajani, Zhang, et al.) (1)

  • Alpaca · Stanford's March 2023 instruction-tuned LLaMA-7B, built for under $600 of synthetic data; proved frontier-adjacent assistants could be replicated for pocket change.

Stanford NLP Group (Jeffrey Pennington, Richard Socher, Christopher Manning) (1)

  • GloVe · Stanford's 2014 word-embedding method that fits vectors to global co-occurrence statistics, the main pre-Transformer alternative to word2vec.

StepFun (1)

  • Step-2 · StepFun's 2024 trillion-parameter-scale mixture-of-experts language model, one of the first Chinese models announced at that scale.

Technology Innovation Institute (1)

  • Falcon (models) · The UAE Technology Innovation Institute's open model line; Falcon 180B was 2023's largest open release, and Falcon Mamba pioneered open SSM scale.

Xiaomi (1)

  • MiMo · Xiaomi's open model line: reasoning-first small models (MiMo-7B, 2025) scaled to the frontier checkpoints that led OpenRouter usage in 2026.

Model families (13)

  • Claude (model family) · Anthropic's assistant model line, from Claude 1 (2023) through the Claude 4.x series to the Fable/Mythos 5 generation (2026).
  • DeepSeek (model family) · DeepSeek's open-weight model line, known for efficiency innovations (MLA, DeepSeekMoE) and the R1 reasoning model.
  • Gemini (model family) · Google DeepMind's flagship multimodal model line, successor to LaMDA and PaLM; natively multimodal since Gemini 1.0 (2023).
  • GLM (model family) · Zhipu AI's model line, from the GLM-130B open release of 2022 through the agentic GLM-4.5/5 generation.
  • GPT (model family) · OpenAI's Generative Pre-trained Transformer series, from GPT-1 (2018) through GPT-4 and successors.
  • Grok (model family) · xAI's model line, distributed through the X platform; known for rapid scaling on the Colossus cluster and a lighter-refusal posture.
  • Kimi (model family) · Moonshot AI's model line: the long-context Kimi assistant, k1.5 reasoning, and the trillion-parameter open-weight K2 series.
  • Llama (model family) · Meta's open-weight model line; the LLaMA leak of 2023 seeded the modern open-source LLM ecosystem.
  • Mistral (model family) · Mistral AI's model line: efficient open-weight releases (Mistral 7B, Mixtral) alongside commercial Large and specialist models.
  • Moondream (model family) · The tiny-VLM line from M87 Labs: Moondream 1 (2024) through the MoE Moondream 3 Preview, built for edge and high-volume vision work.
  • Phi (model family) · Microsoft's small-model line built on the 'textbooks are all you need' thesis: curated and synthetic data over parameter count.
  • Qwen (model family) · Alibaba's open-weight model line; by 2025 the most-downloaded and most-fine-tuned base family in the open ecosystem.
  • Trinity (model family) · Arcee AI's open-weight sparse mixture-of-experts line: edge-scale Nano, agent-focused Mini, the 400B Trinity Large, and a Thinking reasoning branch.

Organizations (47)

  • 01.AI · Kai-Fu Lee's 2023 model startup; its Yi-34B briefly led open leaderboards before the company pivoted away from frontier pretraining.
  • Ai2 · Seattle nonprofit AI research institute (AI2); publisher of the fully open OLMo models, Dolma corpus, and Tulu post-training recipes.
  • AI21 Labs · Israeli AI lab behind the Jurassic dense models and the Jamba attention-SSM hybrids, later pivoted from consumer writing tools to enterprise AI systems.
  • Aleph Alpha · German AI company in Heidelberg that built the Luminous models, pivoted to a sovereign enterprise stack, and agreed to merge with Cohere in 2026.
  • Ant Group · Alipay's parent company; its Ling/Bailing model line pioneered frontier MoE training on mixed domestic silicon.
  • Anthropic · AI safety and research company founded in 2021 by former OpenAI researchers; developer of the Claude model family.
  • Arcee AI · US open intelligence lab that grew from model merging and distillation into from-scratch pretraining, publisher of the Trinity model family.
  • Baichuan · Wang Xiaochuan's 2023 lab; its early open Baichuan models seeded China's ecosystem before a pivot toward medical AI.
  • Baidu · Chinese search giant and AI pioneer; developer of the ERNIE model line and the Kunlun AI chips.
  • ByteDance · TikTok's parent and one of the world's largest AI compute buyers; its Seed lab builds the Doubao assistant and Seedream/Seedance media models.
  • Cerebras Systems · American wafer-scale AI hardware company behind the WSE chips, Cerebras-GPT, the Condor Galaxy supercomputers with G42, and a high-speed inference service.
  • Cohere · Toronto-based enterprise AI lab founded in 2019 by Transformer co-author Aidan Gomez; builds the Command model line, the North agent platform, and the Aya open-science program.
  • Cursor · AI code editor built by Anysphere, one of the largest commercial deployers of frontier coding models and, since late 2025, trainer of its own Composer coding model.
  • Databricks · American data and AI platform company founded by the creators of Apache Spark, acquirer of MosaicML and publisher of the open-weight DBRX model.
  • DeepSeek · Hangzhou AI lab funded by hedge fund High-Flyer; publisher of the open-weight DeepSeek model family and the R1 reasoning model.
  • EleutherAI · Grassroots research collective turned nonprofit; produced GPT-J, GPT-NeoX, The Pile, and the Pythia interpretability suite.
  • Google DeepMind · Google's AI research division, formed in 2023 by merging DeepMind and Google Brain; developer of the Gemini model family.
  • Huawei · China's sanctioned hardware champion: Ascend AI chips, the CloudMatrix rack systems, and the PanGu model line.
  • Hugging Face · The hub of the open machine-learning ecosystem: model hosting, the transformers library, datasets, and leaderboards.
  • iFlytek · Chinese speech-technology giant; its Spark models are notable for training entirely on Huawei Ascend domestic chips.
  • Liquid AI · MIT CSAIL spinoff founded in 2023 by the liquid-neural-network researchers; builds the efficiency-first LFM model line for edge and on-device use, backed by a $250M AMD-led round.
  • Meituan · China's local-services giant; its LongCat models reached 1.6T parameters trained entirely on Chinese chips, leading open agentic-coding rankings in 2026.
  • Meta AI · Meta's AI research and product organization (FAIR, then GenAI, then Superintelligence Labs), developer of the Llama model family.
  • Microsoft AI · Microsoft's AI program and consumer AI division (formed March 2024 under Mustafa Suleyman): OpenAI's largest backer, maker of Copilot, Phi, and the MAI models.
  • MiniMax · Shanghai AI company known for linear-attention frontier models (MiniMax-01, M1/M2), the Hailuo video line, and companion apps.
  • Mistral AI · French AI lab founded in 2023; Europe's most prominent frontier-model company and a leading open-weight publisher.
  • Moondream · M87 Labs' tiny vision-language model line; the reference small VLM for edge and high-volume deployment.
  • Moonshot AI · Beijing AI startup founded by Yang Zhilin in 2023; developer of the Kimi models, including the trillion-parameter open-weight Kimi K2.
  • MosaicML · San Francisco startup (2021-2023) whose efficient-training platform and open-weight MPT models led to a $1.3 billion Databricks acquisition; its team became Databricks Mosaic Research and built DBRX.
  • Ndea · AI research lab founded in January 2025 by Keras creator François Chollet and Zapier co-founder Mike Knoop, pursuing deep-learning-guided program synthesis as a path to AGI.
  • Nous Research · US open-source AI collective turned company; builds the Hermes instruction-tuned model line, the YaRN context-extension method, and the DisTrO/Psyche decentralized-training stack.
  • NVIDIA · The dominant supplier of AI training and inference hardware; its datacenter GPUs are the substrate of the LLM era.
  • OpenAI · San Francisco AI lab founded in 2015; developer of the GPT model series and ChatGPT.
  • Prometheus · Jeff Bezos and Vik Bajaj's AI company, founded November 2025; pursues an 'artificial general engineer' for the physical economy at a reported $41 billion valuation.
  • Qwen team · Alibaba's model group (Tongyi lab); publisher of the Qwen family, the open ecosystem's most-downloaded base-model line.
  • Safe Superintelligence · AI lab founded in June 2024 by Ilya Sutskever, Daniel Gross, and Daniel Levy to build a safe superintelligence as its first and only product.
  • SenseTime · Computer-vision incumbent turned model lab; the SenseNova line runs on one of China's largest private AI datacenters.
  • Shanghai AI Laboratory · State-backed research institute behind the InternLM and InternVL open model lines and the OpenCompass evaluation platform.
  • Stability AI · Company behind the Stable Diffusion image models; emblem of the 2022 open generative-AI wave and its business-model turbulence.
  • StepFun · Shanghai lab founded by ex-Microsoft researcher Jiang Daxin; trillion-parameter Step-2 and the multimodal Step series.
  • Technology Innovation Institute · Abu Dhabi state applied-research institute under the ATRC; pretrained the open Falcon model line on RefinedWeb, including Falcon 180B.
  • Tencent · Chinese internet giant behind WeChat; developer of the Hunyuan model line, including open-weight text, video, and 3D models.
  • Thinking Machines Lab · AI lab founded in February 2025 by former OpenAI CTO Mira Murati; raised a record ~$2B seed at a $12B valuation and builds the Tinker fine-tuning API around a customization-first thesis.
  • xAI · Elon Musk's AI company, founded 2023; developer of the Grok model family and the Colossus supercomputer.
  • Xiaomi · Consumer-electronics giant whose MiMo model line vaulted to open-weight usage leadership in 2026 on aggressive price-performance.
  • Yandex · Russian internet company behind the dominant Russian search engine and Alice assistant; released YaLM-100B openly in 2022 and builds the YandexGPT line.
  • Zhipu AI · Beijing AI company (branded Z.ai) spun out of Tsinghua University; developer of the GLM model family.

Concepts and techniques (60)

  • Activation steering · Adding vectors to a transformer's residual stream at inference time to shift model behavior; used both as a control knob and as causal evidence in interpretability.
  • Activation verbalizer · Anthropic technique in which a language model is trained to translate another model's token-level activations into natural-language summaries.
  • Adam (optimizer) · The adaptive optimization algorithm introduced by Kingma and Ba in 2014, combining momentum with per-parameter learning rates, used to train nearly every modern language model.
  • Analytic distillation · The extraction of what a trained model knows into smaller, cheaper, or more explicit form, spanning classic knowledge distillation, rationale distillation, and mechanistic interpretability.
  • Attention sink · The tendency of autoregressive transformers to concentrate attention mass on initial tokens; exploited by StreamingLLM and built into gpt-oss as learned biases.
  • Backpropagation · The chain-rule algorithm that computes a neural network's loss gradient in a single backward pass, developed by Werbos in the early 1970s and popularized by Rumelhart, Hinton, and Williams in 1986.
  • Boltzmann machine · The stochastic recurrent network of Ackley, Hinton, and Sejnowski (1985) that learns distributions by contrasting clamped and free-running statistics; its restricted form seeded deep belief networks.
  • Chain-of-thought prompting · Eliciting step-by-step reasoning before the answer; discovered as a prompting trick in 2022, later internalized by trained reasoning models.
  • Circuits (interpretability) · Subnetworks of a neural network that implement identifiable algorithms; the central object of study in mechanistic interpretability.
  • Compositional generalization · The capacity to understand and produce novel combinations of known components, the systematicity property long contested in neural networks and a central open question for large language models.
  • Constitutional AI · Anthropic's post-training method: models critique and revise their own outputs against written principles, then train on AI-generated preferences.
  • Cross-entropy loss · The standard training loss of language modeling: the negative log probability assigned to the correct token, the quantity plotted on every scaling-law curve.
  • Diffusion language model · Text generation by iterative parallel denoising of masked or corrupted token sequences, an alternative to left-to-right autoregression.
  • Direct preference optimization · Preference-alignment method that optimizes a policy directly on comparison data via a closed-form objective, skipping the reward model and RL loop.
  • Distributed training · Techniques for training models across thousands of accelerators: data, tensor, pipeline, and expert parallelism, plus sharding schemes like ZeRO/FSDP.
  • Dropout · The regularization technique that randomly zeroes hidden units during training, standard in early Transformers and largely retired as pretraining shifted to single-epoch, web-scale data.
  • Eliciting Latent Knowledge · Open alignment problem posed by ARC in 2021: training a model to report what it internally knows rather than what looks correct to human overseers.
  • Embedding (machine learning) · The mapping from discrete tokens to continuous vectors that forms the input and output interface of neural language models and encodes semantic similarity as geometry.
  • Feed-forward network · The position-wise two-layer perceptron inside every Transformer block, holding most of the stack's parameters and complementing attention's routing with per-token computation.
  • FlashAttention · IO-aware exact attention algorithm that tiles the computation to fit in GPU on-chip memory; the de facto standard attention kernel in training and serving stacks.
  • Hopfield network · John Hopfield's 1982 recurrent neural network that stores patterns as minima of an Ising-like energy function and retrieves them by relaxation, the model that made associative memory rigorous.
  • Hybrid architecture (LLM) · Model designs that interleave self-attention with state-space, recurrent, or linear-attention layers, or mix sliding-window and full attention, to cut KV-cache and compute costs.
  • In-context learning · The ability of large language models to perform a new task from instructions or examples in the prompt alone, without any weight update, demonstrated at scale by GPT-3 in 2020.
  • Instruction tuning · Supervised fine-tuning of pretrained language models on instruction-response pairs, teaching them to follow natural-language tasks zero-shot.
  • Keras · Open-source deep learning API created by François Chollet in March 2015, rewritten in 2023 to run on JAX, TensorFlow, and PyTorch, and used as launch tooling for Gemma and other open LLMs.
  • Knowledge distillation · Training a smaller student model to imitate a larger teacher; the mechanism behind small frontier-adjacent models and a persistent IP controversy.
  • Large language model · A neural language model, typically a decoder-only Transformer with billions of parameters, pretrained on internet-scale text.
  • Layer normalization · Per-token normalization of activations across the feature dimension, introduced by Ba, Kiros, and Hinton in 2016 and built into every Transformer block as part of the add-and-norm pattern.
  • Logit lens · Decoding a transformer's intermediate residual-stream states through its output unembedding, revealing layer-by-layer refinement of next-token predictions.
  • LSTM · Long short-term memory: the 1997 recurrent architecture that dominated sequence modeling for two decades before the Transformer.
  • Masked language modeling · The pretraining objective of predicting hidden tokens from bidirectional context, introduced at scale by BERT in 2018 and the training signal of the encoder-only lineage.
  • Mechanistic interpretability · The research program of reverse-engineering neural networks into human-understandable features and circuits, from the 2020 Distill circuits thread to routine white-box analyses in frontier system cards.
  • Mixture of experts · Sparse architecture in which a router activates only a few expert subnetworks per token, decoupling parameter count from per-token compute.
  • Multi-head attention · Running attention in parallel subspaces within each layer; its head-sharing variants (MQA, GQA, MLA) set the size of the KV cache and much of the cost of inference.
  • Multi-head latent attention · Attention variant that caches a low-rank latent compression of keys and values instead of full per-head tensors; introduced in DeepSeek-V2 (2024).
  • Next-token prediction · The objective of predicting each token from all preceding ones, the training signal behind GPT-style language models and the foundation of autoregressive generation.
  • PagedAttention · Virtual-memory-style paging for the KV cache, introduced with vLLM in 2023; cut serving memory waste to near zero and became standard across inference stacks.
  • Perceptron · Frank Rosenblatt's 1957 learning machine: the first artificial neural network shown to learn its weights from data, with a provably convergent rule and a famous linear-separability limitation.
  • Positional encoding · The mechanism that injects token order into the otherwise permutation-invariant Transformer, evolving from fixed sinusoidal vectors to the rotary embeddings used by most modern LLMs.
  • Pretraining · The first phase of language model development: self-supervised training over large text corpora that produces a base model, before post-training adapts it for use.
  • Reinforcement learning from human feedback · Post-training method that optimizes a language model against a reward model trained on human preference comparisons; the technique behind ChatGPT-style assistants.
  • Reinforcement learning with verifiable rewards · Post-training with programmatic reward signals (test suites, answer checkers) instead of learned reward models; the engine of the reasoning-model era.
  • Reward model · A model trained to score another model's outputs against human or AI preferences, supplying the training signal for RLHF and, in process form, for reasoning search.
  • RMSNorm · Root mean square layer normalization: a simplification of layer normalization that rescales activations by their root mean square alone, the default normalizer of modern open-weight LLMs.
  • Scaling laws · Empirical power-law relationships between model performance and parameters, data, and compute; the planning instrument of the LLM era.
  • Self-attention · Mechanism by which each position in a sequence computes its representation as an attention-weighted sum over all positions; the core operation of the Transformer.
  • Seq2seq · Sequence-to-sequence learning: the 2014 encoder-decoder paradigm that mapped variable-length input to output through a fixed context vector, the direct ancestor of the attention era.
  • Sliding-window attention · Restricting each token's attention to a fixed-size local window, cutting attention cost from quadratic to linear in context length and bounding KV-cache memory.
  • Softmax · The function that turns a vector of raw scores into a probability distribution, used in attention weighting, vocabulary output layers, and mixture-of-experts routing.
  • Sparse attention · Attention variants in which each token attends to a selected subset of positions instead of the full sequence, cutting the quadratic cost of self-attention.
  • Sparse autoencoder · Dictionary learning on language model activations; decomposes superposed representations into sparse, often monosemantic features.
  • State-space model · Sequence architecture family (S4, Mamba) that replaces attention with a fixed-size recurrent state, giving linear-time processing and constant-memory generation.
  • Superposition (interpretability) · The hypothesis that neural networks represent more features than they have neurons by packing them into overlapping directions, explaining polysemantic neurons and motivating sparse autoencoders.
  • Supervised fine-tuning · Post-training method that teaches a pretrained language model desired behavior by supervised next-token training on curated prompt-response demonstrations.
  • Teacher forcing · The training protocol that feeds a sequence model the true previous tokens instead of its own predictions, enabling fully parallel next-token training at the cost of exposure bias.
  • Test-time compute · Improving answers by spending more computation at inference (longer reasoning chains, search, sampling) rather than more pretraining; the axis behind reasoning models.
  • Tokenization · The conversion of raw text into the integer token ids that language models read and write, dominated by byte-pair encoding variants with vocabularies of tens to hundreds of thousands of tokens.
  • Transformer (architecture) · Neural network architecture based on self-attention, introduced by Vaswani et al. in 2017; the basis of virtually all modern large language models.
  • Vision Transformer · Transformer applied to images by treating fixed-size patches as tokens; introduced by Dosovitskiy et al. in 2020, now the standard vision encoder in multimodal LLMs.
  • Word2vec · The 2013 word-embedding method whose vector arithmetic ('king - man + woman = queen') made distributed representations the foundation of neural NLP.

Benchmarks and datasets (5)

  • Common Crawl · A nonprofit's free archive of web crawls, petabytes of raw page data refreshed monthly, that serves as the raw substrate for most large language model training corpora.
  • Common Pile · EleutherAI's June 2025 corpus of 8 TB of public-domain and openly licensed text from 30 sources, released with the Comma v0.1 models as an openly licensed successor to the Pile.
  • ImageNet · The 14-million-image, WordNet-organized labeled dataset (2009) whose ILSVRC challenge (2010-2017) was the benchmark of the deep learning revolution, from AlexNet to ResNet.
  • MMLU · Massive Multitask Language Understanding, a 57-subject multiple-choice knowledge benchmark introduced in 2020 and long the standard headline eval for language models.
  • The Pile · EleutherAI's December 2020 corpus of 825 GiB of curated English text drawn from 22 sources, the training set behind GPT-Neo, GPT-J, GPT-NeoX-20B, and Pythia, later partly withdrawn amid the Books3 copyright dispute.

Incidents and governance (2)

People (20)

  • Alex Krizhevsky · Canadian computer scientist who wrote and trained AlexNet, the GPU-based convolutional network whose 2012 ImageNet victory launched the deep learning era, and created the CIFAR-10 and CIFAR-100 datasets.
  • Andrej Karpathy · Slovak-Canadian computer scientist; co-founder of OpenAI, director of AI at Tesla from 2017 to 2022, creator of Stanford's CS231n course and of educational artifacts including nanoGPT, and a pretraining research lead at Anthropic since 2026.
  • Andrew Ng · British-American computer scientist who co-founded Google Brain and Coursera, served as Baidu's chief scientist from 2014 to 2017, and built online machine learning courses taken by millions of students.
  • Arthur Samuel · American computer pioneer whose IBM checkers program (1952-1967) was the first celebrated self-improving machine and whose 1959 paper gave machine learning its name.
  • David Rumelhart · American psychologist who led the 1980s connectionist revival through the PDP volumes and the 1986 Nature paper that popularized backpropagation; his past-tense model ignited the rules-versus-associations debate.
  • David Silver · British computer scientist who led the AlphaGo and AlphaZero projects at DeepMind; 2019 ACM Prize in Computing laureate and founder of Ineffable Intelligence.
  • Demis Hassabis · British AI researcher and entrepreneur; co-founder and CEO of Google DeepMind and co-recipient of the 2024 Nobel Prize in Chemistry for AlphaFold's protein structure prediction.
  • Frank Rosenblatt · American psychologist who invented the perceptron, the first neural network trained by a provably convergent learning rule, at Cornell Aeronautical Laboratory in 1957.
  • François Chollet · French software engineer and AI researcher, creator of Keras and the Xception architecture, author of On the Measure of Intelligence and the ARC-AGI benchmark, and co-founder of the AGI lab Ndea.
  • Geoffrey Hinton · British-Canadian computer scientist who popularized backpropagation, co-invented the Boltzmann machine, led the deep learning revival behind AlexNet, won the 2018 Turing Award and the 2024 Nobel Prize in Physics, and left Google in 2023 to warn about AI risk.
  • Ian Goodfellow · American computer scientist who invented the generative adversarial network, co-authored the Deep Learning textbook, and held research roles at Google Brain, OpenAI, Apple, and Google DeepMind.
  • Ilya Sutskever · Israeli-Canadian computer scientist who co-created AlexNet and sequence-to-sequence learning, co-founded OpenAI and served as its chief scientist through the GPT era, and co-founded Safe Superintelligence in 2024.
  • John Hopfield · American physicist whose 1982 associative-memory model, the Hopfield network, helped restart neural network research; co-recipient of the 2024 Nobel Prize in Physics.
  • John Jumper · American chemist and computer scientist who led the AlphaFold project at Google DeepMind; co-recipient of the 2024 Nobel Prize in Chemistry for protein structure prediction.
  • Paul Werbos · American researcher who first described backpropagation in his 1974 Harvard dissertation, pioneered recurrent backpropagation and adaptive critic designs, and directed neural network funding at NSF from 1988 to 2015.
  • Richard Sutton · Canadian computer scientist who built the conceptual and algorithmic foundations of reinforcement learning and won the 2024 ACM Turing Award with Andrew Barto.
  • Walter Pitts · American logician who co-created the first mathematical model of the neuron in 1943 with Warren McCulloch, founding the theory of neural networks.
  • Warren McCulloch · American neurophysiologist and psychiatrist who co-founded neural network theory with the 1943 McCulloch-Pitts paper and chaired the Macy conferences on cybernetics.
  • Yann LeCun · French-American computer scientist who invented convolutional neural networks for document recognition, founded and led Meta's FAIR laboratory as chief AI scientist, co-received the 2018 Turing Award, and left Meta in 2025 to found the world-model startup AMI Labs.
  • Yoshua Bengio · Canadian computer scientist at the Université de Montréal, founder of Mila, pioneer of neural language models and attention-based machine translation, co-recipient of the 2018 Turing Award, and since 2023 a leading scientific voice on AI catastrophic risk.

Reference (3)

  • Chinese AI labs · The complete landscape of Chinese organizations with significant model-pretraining capacity: frontier labs, tech giants, state institutes, and the domestic-silicon question.
  • Seminal AI essays · A reading guide to the thirteen essays and blog posts that most shaped how the field understands large language models, scaling, and AI trajectories.
  • Stated missions of AI labs · Every major lab's stated mission in its own words, beside the work each is known for; sources and record analysis on each organization page.