[ { "name": "Transformer", "family": null, "org": "Google Brain / Google Research", "release_date": "2017-06-12", "paper_arxiv_id": "1706.03762", "paper_url": "https://arxiv.org/abs/1706.03762", "params": "65M (base) / 213M (big)", "weights": "open-research-only", "architecture": "encoder-decoder transformer", "notes": "The 'Attention Is All You Need' paper by Vaswani et al.; introduced multi-head self-attention and the encoder-decoder transformer architecture that became the substrate of every subsequent LLM." }, { "name": "ULMFiT", "family": null, "org": "fast.ai / Aylien", "release_date": "2018-01-18", "paper_arxiv_id": "1801.06146", "paper_url": "https://arxiv.org/abs/1801.06146", "params": "~24M (AWD-LSTM)", "weights": "open", "architecture": "AWD-LSTM (3-layer)", "notes": "Howard & Ruder popularized the pretrain-then-fine-tune recipe for NLP; pre-transformer but methodologically foundational for the LLM era." }, { "name": "ELMo", "family": null, "org": "Allen Institute for AI / U. Washington", "release_date": "2018-02-15", "paper_arxiv_id": "1802.05365", "paper_url": "https://arxiv.org/abs/1802.05365", "params": "~94M", "weights": "open", "architecture": "bidirectional LSTM", "notes": "Deep contextualized word representations from a biLSTM language model; first to popularize contextual embeddings that varied with usage." }, { "name": "GPT-1", "family": "GPT", "org": "OpenAI", "release_date": "2018-06-11", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/language-unsupervised/", "params": "117M", "weights": "open", "architecture": "decoder-only transformer (12 layers)", "notes": "'Improving Language Understanding by Generative Pre-Training' (Radford et al.); first generative-pretrained transformer LM; published as OpenAI tech report, never on arXiv." }, { "name": "BERT", "family": "BERT", "org": "Google AI Language", "release_date": "2018-10-11", "paper_arxiv_id": "1810.04805", "paper_url": "https://arxiv.org/abs/1810.04805", "params": "110M (base) / 340M (large)", "weights": "open", "architecture": "encoder-only transformer", "notes": "Bidirectional Encoder Representations from Transformers; masked-LM + next-sentence-prediction pretraining; the inflection paper that made pretrain-then-fine-tune NLP's dominant paradigm." }, { "name": "GPT-2", "family": "GPT", "org": "OpenAI", "release_date": "2019-02-14", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/better-language-models/", "params": "1.5B (full); 117M/345M/762M (smaller)", "weights": "open", "architecture": "decoder-only transformer", "notes": "Initial announcement Feb 14 2019 with staged release; full 1.5B model released Nov 5 2019. First high-profile 'too dangerous to release' moment; never an arXiv submission, only OpenAI's blog + tech report." }, { "name": "XLNet", "family": null, "org": "Google Brain / CMU", "release_date": "2019-06-19", "paper_arxiv_id": "1906.08237", "paper_url": "https://arxiv.org/abs/1906.08237", "params": "340M (large)", "weights": "open", "architecture": "permutation-LM transformer (Transformer-XL backbone)", "notes": "Permutation language modeling objective; briefly displaced BERT atop GLUE in mid-2019." }, { "name": "RoBERTa", "family": "BERT", "org": "Facebook AI Research (Meta)", "release_date": "2019-07-26", "paper_arxiv_id": "1907.11692", "paper_url": "https://arxiv.org/abs/1907.11692", "params": "125M (base) / 355M (large)", "weights": "open", "architecture": "encoder-only transformer", "notes": "'Robustly Optimized BERT'; showed BERT was undertrained; longer, larger-batch pretraining without next-sentence prediction; established that scale and data trump objective tweaks." }, { "name": "T5", "family": "T5", "org": "Google Research", "release_date": "2019-10-23", "paper_arxiv_id": "1910.10683", "paper_url": "https://arxiv.org/abs/1910.10683", "params": "60M / 220M / 770M / 3B / 11B", "weights": "open", "architecture": "encoder-decoder transformer", "notes": "Text-to-Text Transfer Transformer; cast every NLP task as text-to-text; the 11B version was, at release, the largest open-released encoder-decoder LM." }, { "name": "BART", "family": null, "org": "Facebook AI Research (Meta)", "release_date": "2019-10-29", "paper_arxiv_id": "1910.13461", "paper_url": "https://arxiv.org/abs/1910.13461", "params": "140M (base) / 400M (large)", "weights": "open", "architecture": "encoder-decoder transformer (denoising autoencoder)", "notes": "Bidirectional encoder + autoregressive decoder pretrained as a denoising autoencoder; strong on summarization and generation." }, { "name": "GPT-3", "family": "GPT", "org": "OpenAI", "release_date": "2020-05-28", "paper_arxiv_id": "2005.14165", "paper_url": "https://arxiv.org/abs/2005.14165", "params": "175B (plus 13 smaller sizes)", "weights": "closed-api", "architecture": "decoder-only transformer (96 layers)", "notes": "'Language Models are Few-Shot Learners'; the defining 175B model that introduced in-context learning at scale and launched the modern LLM era." }, { "name": "mT5", "family": "T5", "org": "Google Research", "release_date": "2020-10-22", "paper_arxiv_id": "2010.11934", "paper_url": "https://arxiv.org/abs/2010.11934", "params": "300M to 13B", "weights": "open", "architecture": "encoder-decoder transformer", "notes": "Multilingual T5 trained on mC4 covering 101 languages; benchmark multilingual pretrained model of the era." }, { "name": "WuDao 2.0", "family": "WuDao", "org": "Beijing Academy of Artificial Intelligence (BAAI)", "release_date": "2021-06-01", "paper_arxiv_id": null, "paper_url": "https://en.wikipedia.org/wiki/Wu_Dao", "params": "1.75T", "weights": "unknown", "architecture": "MoE multimodal transformer", "notes": "Chinese state-supported trillion-parameter multimodal model announced at the 2021 BAAI conference; weights never broadly released; primarily a scaling milestone." }, { "name": "Codex", "family": "GPT", "org": "OpenAI", "release_date": "2021-07-07", "paper_arxiv_id": "2107.03374", "paper_url": "https://arxiv.org/abs/2107.03374", "params": "12B (largest evaluated); GPT-3 family", "weights": "closed-api", "architecture": "decoder-only transformer", "notes": "GPT-3 fine-tuned on GitHub; powered the first GitHub Copilot; introduced the HumanEval benchmark." }, { "name": "Megatron-Turing NLG 530B", "family": null, "org": "Microsoft / NVIDIA", "release_date": "2021-10-11", "paper_arxiv_id": "2201.11990", "paper_url": "https://arxiv.org/abs/2201.11990", "params": "530B", "weights": "closed-api", "architecture": "decoder-only transformer (dense)", "notes": "Joint Microsoft/NVIDIA model; largest dense LM at announcement; 3x GPT-3 by params. Paper Jan 2022 (arxiv 2201.11990); model announced Oct 11 2021." }, { "name": "Gopher", "family": null, "org": "DeepMind", "release_date": "2021-12-08", "paper_arxiv_id": "2112.11446", "paper_url": "https://arxiv.org/abs/2112.11446", "params": "280B (also 44M, 117M, 417M, 1.4B, 7.1B)", "weights": "closed-api", "architecture": "decoder-only transformer", "notes": "DeepMind's first frontier-scale LLM; evaluated across 152 tasks; main vehicle for studying scaling effects pre-Chinchilla." }, { "name": "LaMDA", "family": null, "org": "Google Research", "release_date": "2022-01-20", "paper_arxiv_id": "2201.08239", "paper_url": "https://arxiv.org/abs/2201.08239", "params": "137B", "weights": "closed-api", "architecture": "decoder-only transformer", "notes": "Dialog-specialized LM; foundation of early Bard. Initially announced May 2021 at Google I/O; paper landed Jan 2022." }, { "name": "Chinchilla", "family": null, "org": "DeepMind", "release_date": "2022-03-29", "paper_arxiv_id": "2203.15556", "paper_url": "https://arxiv.org/abs/2203.15556", "params": "70B", "weights": "closed-api", "architecture": "decoder-only transformer", "notes": "Compute-optimal scaling laws: model and tokens should scale 1:1; 70B model with 4x more data outperformed 280B Gopher; reshaped how the field budgets compute." }, { "name": "PaLM", "family": "PaLM", "org": "Google Research", "release_date": "2022-04-05", "paper_arxiv_id": "2204.02311", "paper_url": "https://arxiv.org/abs/2204.02311", "params": "540B (also 8B, 62B)", "weights": "closed-api", "architecture": "decoder-only transformer (dense)", "notes": "Pathways Language Model; densely activated 540B trained on TPU v4 pods; first dense LM to convincingly outperform humans on BIG-Bench." }, { "name": "OPT-175B", "family": "OPT", "org": "Meta AI", "release_date": "2022-05-02", "paper_arxiv_id": "2205.01068", "paper_url": "https://arxiv.org/abs/2205.01068", "params": "125M to 175B", "weights": "open-research-only", "architecture": "decoder-only transformer", "notes": "First GPT-3-scale open model from a major lab; trained at 1/7 the GPT-3 carbon footprint; released with a full training logbook." }, { "name": "BLOOM", "family": null, "org": "BigScience (Hugging Face-led collaboration)", "release_date": "2022-07-11", "paper_arxiv_id": "2211.05100", "paper_url": "https://arxiv.org/abs/2211.05100", "params": "176B", "weights": "open", "architecture": "decoder-only transformer", "notes": "First fully open 175B-scale multilingual LM (46 natural + 13 programming languages); developed by ~1000 researchers; released under RAIL license. Weights released July 11 2022; arXiv paper Nov 2022." }, { "name": "Flan-T5 / Flan-PaLM", "family": "T5 / PaLM", "org": "Google Research", "release_date": "2022-10-20", "paper_arxiv_id": "2210.11416", "paper_url": "https://arxiv.org/abs/2210.11416", "params": "T5: 80M-11B; PaLM: 540B", "weights": "open (Flan-T5) / closed-api (Flan-PaLM)", "architecture": "encoder-decoder / decoder-only transformer", "notes": "'Scaling Instruction-Finetuned Language Models'; finetuned on 1.8K tasks with chain-of-thought; the prototype instruction-tuning recipe that everyone copied." }, { "name": "Galactica", "family": null, "org": "Meta AI", "release_date": "2022-11-16", "paper_arxiv_id": "2211.09085", "paper_url": "https://arxiv.org/abs/2211.09085", "params": "125M / 1.3B / 6.7B / 30B / 120B", "weights": "open-research-only", "architecture": "decoder-only transformer", "notes": "Trained on 106B tokens of scientific text (papers, knowledge bases); demo pulled within 3 days after hallucinated-citations backlash, marking the era when public demos started getting scrutinized." }, { "name": "ChatGPT", "family": "GPT", "org": "OpenAI", "release_date": "2022-11-30", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/chatgpt/", "params": "GPT-3.5 lineage (unknown exact)", "weights": "closed-api", "architecture": "decoder-only transformer (RLHF fine-tune of text-davinci-002/003)", "notes": "Consumer-product breakout; 1M users in 5 days, 100M in 2 months; fastest-growing consumer technology product in history at the time; the moment LLMs became mainstream." }, { "name": "LLaMA-1", "family": "LLaMA", "org": "Meta AI / FAIR", "release_date": "2023-02-27", "paper_arxiv_id": "2302.13971", "paper_url": "https://arxiv.org/abs/2302.13971", "params": "7B / 13B / 33B / 65B", "weights": "open-research-only", "architecture": "decoder-only transformer (RoPE, SwiGLU, RMSNorm)", "notes": "Compute-optimal trained on 1-1.4T tokens; weights leaked to 4chan March 2023, kick-starting the open-weight LLaMA era; established the modern open-LLM blueprint." }, { "name": "GPT-4", "family": "GPT", "org": "OpenAI", "release_date": "2023-03-14", "paper_arxiv_id": "2303.08774", "paper_url": "https://arxiv.org/abs/2303.08774", "params": "unknown (rumored ~1.8T MoE)", "weights": "closed-api", "architecture": "rumored MoE multimodal transformer", "notes": "GPT-4 Technical Report; first model with human-expert performance on professional/academic benchmarks; OpenAI declined to disclose architecture or training details." }, { "name": "Claude 1", "family": "Claude", "org": "Anthropic", "release_date": "2023-03-14", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/introducing-claude", "params": "unknown", "weights": "closed-api", "architecture": "decoder-only transformer (Constitutional AI-trained)", "notes": "First Claude release; available only via API to selected partners; introduced Constitutional AI alignment approach." }, { "name": "PaLM 2", "family": "PaLM", "org": "Google", "release_date": "2023-05-10", "paper_arxiv_id": "2305.10403", "paper_url": "https://arxiv.org/abs/2305.10403", "params": "unknown (340B rumored)", "weights": "closed-api", "architecture": "decoder-only transformer", "notes": "Backbone of Bard and Google's Duet AI; emphasized multilingual + reasoning capability with smaller compute; announced at Google I/O 2023." }, { "name": "Falcon-40B", "family": "Falcon", "org": "Technology Innovation Institute (TII), Abu Dhabi", "release_date": "2023-05-25", "paper_arxiv_id": "2311.16867", "paper_url": "https://arxiv.org/abs/2311.16867", "params": "40B (also 7B)", "weights": "open", "architecture": "decoder-only transformer (multi-query attention)", "notes": "First major UAE-origin open LLM; trained on the RefinedWeb dataset; topped HuggingFace's Open LLM Leaderboard for several months in mid-2023." }, { "name": "Claude 2", "family": "Claude", "org": "Anthropic", "release_date": "2023-07-11", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-2", "params": "unknown", "weights": "closed-api", "architecture": "decoder-only transformer", "notes": "First Anthropic model available to the general public via claude.ai; 100K-token context window; surge of long-document use cases." }, { "name": "LLaMA-2", "family": "LLaMA", "org": "Meta AI", "release_date": "2023-07-18", "paper_arxiv_id": "2307.09288", "paper_url": "https://arxiv.org/abs/2307.09288", "params": "7B / 13B / 70B", "weights": "open", "architecture": "decoder-only transformer (GQA on 70B)", "notes": "First Meta LLM with a permissive commercial license; pretrained on 2T tokens; Chat variants RLHF-fine-tuned. The release that made open LLMs commercially viable." }, { "name": "Falcon-180B", "family": "Falcon", "org": "Technology Innovation Institute (TII), Abu Dhabi", "release_date": "2023-09-06", "paper_arxiv_id": null, "paper_url": "https://www.tii.ae/news/technology-innovation-institute-introduces-worlds-most-powerful-open-llm-falcon-180b", "params": "180B", "weights": "open", "architecture": "decoder-only transformer (multi-query attention)", "notes": "Largest open LLM at release; trained on 3.5T tokens; briefly top of the HF leaderboard before LLaMA-2 + fine-tunes overtook it." }, { "name": "Mistral 7B", "family": "Mistral", "org": "Mistral AI", "release_date": "2023-09-27", "paper_arxiv_id": "2310.06825", "paper_url": "https://arxiv.org/abs/2310.06825", "params": "7.3B", "weights": "open", "architecture": "decoder-only transformer (GQA + sliding-window attention)", "notes": "Apache 2.0 7B model that beat LLaMA-2 13B across the board; effectively redefined what was possible at 7B and made Mistral AI a major European player." }, { "name": "Qwen-1.0", "family": "Qwen", "org": "Alibaba Cloud", "release_date": "2023-09-28", "paper_arxiv_id": "2309.16609", "paper_url": "https://arxiv.org/abs/2309.16609", "params": "1.8B / 7B / 14B / 72B", "weights": "open", "architecture": "decoder-only transformer", "notes": "First Alibaba open LLM family (rebranded from Tongyi Qianwen); bilingual EN/CN, pretrained on up to 3T tokens; start of the Qwen series that became China's flagship open-weight lineage." }, { "name": "Yi-34B", "family": "Yi", "org": "01.AI", "release_date": "2023-11-02", "paper_arxiv_id": "2403.04652", "paper_url": "https://arxiv.org/abs/2403.04652", "params": "6B / 34B", "weights": "open", "architecture": "decoder-only transformer", "notes": "Bilingual EN/CN models trained on 3.1T tokens; founded by Kai-Fu Lee; Yi-34B briefly held the top spot on HF's leaderboard." }, { "name": "Grok-1", "family": "Grok", "org": "xAI", "release_date": "2023-11-04", "paper_arxiv_id": null, "paper_url": "https://x.ai/news/grok-os", "params": "314B (MoE; ~86B active)", "weights": "open (weights released March 17 2024)", "architecture": "MoE transformer", "notes": "xAI's debut chatbot; trained in 2 months; weights open-sourced under Apache 2.0 in March 2024; integrated with X (Twitter)." }, { "name": "DeepSeek LLM", "family": "DeepSeek", "org": "DeepSeek AI", "release_date": "2023-11-29", "paper_arxiv_id": "2401.02954", "paper_url": "https://arxiv.org/abs/2401.02954", "params": "7B / 67B", "weights": "open", "architecture": "decoder-only transformer (LLaMA-like)", "notes": "DeepSeek's debut LLM; 67B variant surpassed LLaMA-2 70B on code, math, reasoning; presented its own scaling-laws study. Models released Nov 29 2023; paper Jan 2024." }, { "name": "Gemini 1.0", "family": "Gemini", "org": "Google DeepMind", "release_date": "2023-12-06", "paper_arxiv_id": "2312.11805", "paper_url": "https://arxiv.org/abs/2312.11805", "params": "Ultra / Pro / Nano (sizes undisclosed)", "weights": "closed-api", "architecture": "natively multimodal transformer", "notes": "Google DeepMind's first natively multimodal frontier family; Ultra was the first model to surpass human experts on MMLU at 90.0%." }, { "name": "Mixtral 8x7B", "family": "Mistral", "org": "Mistral AI", "release_date": "2023-12-11", "paper_arxiv_id": "2401.04088", "paper_url": "https://arxiv.org/abs/2401.04088", "params": "47B total / 13B active (8 experts, top-2 routing)", "weights": "open", "architecture": "Sparse Mixture-of-Experts transformer", "notes": "First commercially viable open MoE LLM; matched or beat LLaMA-2 70B and GPT-3.5 across benchmarks; weights released via magnet link Dec 8 2023, paper Jan 8 2024." }, { "name": "Gemini 1.5 Pro", "family": "Gemini", "org": "Google DeepMind", "release_date": "2024-02-15", "paper_arxiv_id": "2403.05530", "paper_url": "https://arxiv.org/abs/2403.05530", "params": "unknown (MoE)", "weights": "closed-api", "architecture": "MoE multimodal transformer", "notes": "Introduced 1M-token context (later 2M); first frontier model with truly long-context multimodal understanding (video, audio, full codebases)." }, { "name": "Gemma 1", "family": "Gemma", "org": "Google DeepMind", "release_date": "2024-02-21", "paper_arxiv_id": "2403.08295", "paper_url": "https://arxiv.org/abs/2403.08295", "params": "2B / 7B", "weights": "open", "architecture": "decoder-only transformer (Gemini-derived)", "notes": "Google's first open-weight model family from the Gemini program; outperformed Mistral 7B and LLaMA-2 across many benchmarks." }, { "name": "Claude 3 (Opus, Sonnet, Haiku)", "family": "Claude", "org": "Anthropic", "release_date": "2024-03-04", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-3-family", "params": "unknown", "weights": "closed-api", "architecture": "multimodal transformer", "notes": "First Claude family with vision and three-tier size structure (Opus/Sonnet/Haiku); Opus briefly matched or beat GPT-4 on most benchmarks." }, { "name": "InternLM 2", "family": "InternLM", "org": "Shanghai AI Lab", "release_date": "2024-03-26", "paper_arxiv_id": "2403.17297", "paper_url": "https://arxiv.org/abs/2403.17297", "params": "1.8B / 7B / 20B", "weights": "open", "architecture": "decoder-only transformer", "notes": "Major Chinese open-weight LM with strong long-context evaluation; from Shanghai AI Lab (SenseTime + universities)." }, { "name": "DBRX", "family": null, "org": "Databricks (Mosaic AI)", "release_date": "2024-03-27", "paper_arxiv_id": null, "paper_url": "https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm", "params": "132B total / 36B active", "weights": "open", "architecture": "fine-grained MoE transformer (16 experts, top-4 routing)", "notes": "Databricks' open LLM; finer-grained MoE than Mixtral; pretrained on 12T tokens; up to 2x more compute-efficient at inference than LLaMA-2 70B." }, { "name": "Llama 3", "family": "LLaMA", "org": "Meta AI", "release_date": "2024-04-18", "paper_arxiv_id": "2407.21783", "paper_url": "https://arxiv.org/abs/2407.21783", "params": "8B / 70B", "weights": "open", "architecture": "decoder-only transformer (GQA, 128K-token tokenizer)", "notes": "Trained on 15T tokens (5x Llama 2); new 128K vocabulary tokenizer; the 70B briefly matched GPT-4 on several benchmarks." }, { "name": "Phi-3", "family": "Phi", "org": "Microsoft Research", "release_date": "2024-04-22", "paper_arxiv_id": "2404.14219", "paper_url": "https://arxiv.org/abs/2404.14219", "params": "3.8B (mini) / 7B (small) / 14B (medium)", "weights": "open", "architecture": "decoder-only transformer", "notes": "Phi-3-mini at 3.8B rivaled Mixtral 8x7B and GPT-3.5; built on the 'Textbooks Are All You Need' data-quality philosophy; first phone-deployable frontier-ish SLM." }, { "name": "DeepSeek-V2", "family": "DeepSeek", "org": "DeepSeek AI", "release_date": "2024-05-07", "paper_arxiv_id": "2405.04434", "paper_url": "https://arxiv.org/abs/2405.04434", "params": "236B total / 21B active", "weights": "open", "architecture": "MoE transformer (Multi-head Latent Attention, DeepSeekMoE)", "notes": "Introduced MLA to slash KV-cache by 93% and triggered an industry-wide API price war (priced at ~1 yuan per million tokens)." }, { "name": "GPT-4o", "family": "GPT", "org": "OpenAI", "release_date": "2024-05-13", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/hello-gpt-4o/", "params": "unknown", "weights": "closed-api", "architecture": "natively multimodal transformer", "notes": "First natively-multimodal GPT (text/audio/image/video in & out); 2x faster and half the price of GPT-4 Turbo; responds to audio in ~232ms (human-conversational latency)." }, { "name": "Falcon 2", "family": "Falcon", "org": "TII Abu Dhabi", "release_date": "2024-05-14", "paper_arxiv_id": "2407.14885", "paper_url": "https://arxiv.org/abs/2407.14885", "params": "11B", "weights": "open", "architecture": "decoder-only transformer (multimodal Falcon 2 VLM variant)", "notes": "Second-generation TII Falcon; 11 languages; 5.5T training tokens; also offered as a vision-language model variant." }, { "name": "Qwen2", "family": "Qwen", "org": "Alibaba Cloud", "release_date": "2024-06-07", "paper_arxiv_id": "2407.10671", "paper_url": "https://arxiv.org/abs/2407.10671", "params": "0.5B / 1.5B / 7B / 57B-A14B (MoE) / 72B", "weights": "open", "architecture": "decoder-only transformer (+ MoE variant)", "notes": "Qwen2-72B matched or beat LLaMA-3 70B and Mistral Large on most benchmarks; 30-language support; weights released first week of June, paper Jul 15." }, { "name": "Claude 3.5 Sonnet", "family": "Claude", "org": "Anthropic", "release_date": "2024-06-20", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-3-5-sonnet", "params": "unknown", "weights": "closed-api", "architecture": "multimodal transformer", "notes": "Mid-tier model outperformed flagship Claude 3 Opus at 1/5 the cost; introduced Artifacts UI; widely regarded as the best coding model of mid-2024." }, { "name": "Gemma 2", "family": "Gemma", "org": "Google DeepMind", "release_date": "2024-06-27", "paper_arxiv_id": "2408.00118", "paper_url": "https://arxiv.org/abs/2408.00118", "params": "2B / 9B / 27B", "weights": "open", "architecture": "decoder-only transformer (interleaved local/global attention, knowledge distillation)", "notes": "27B variant became one of the highest-ranking open LMs on LMSYS Chatbot Arena, outperforming much larger models." }, { "name": "Llama 3.1", "family": "LLaMA", "org": "Meta AI", "release_date": "2024-07-23", "paper_arxiv_id": "2407.21783", "paper_url": "https://arxiv.org/abs/2407.21783", "params": "8B / 70B / 405B", "weights": "open", "architecture": "decoder-only transformer (GQA, 128K context)", "notes": "First frontier-class open model; the 405B matched GPT-4 / Claude 3.5 Sonnet on many benchmarks; 'The Llama 3 Herd of Models' paper introduced the full family." }, { "name": "Mistral Large 2", "family": "Mistral", "org": "Mistral AI", "release_date": "2024-07-24", "paper_arxiv_id": null, "paper_url": "https://mistral.ai/news/mistral-large-2407/", "params": "123B", "weights": "open-research-only", "architecture": "decoder-only transformer (dense)", "notes": "Dense 123B with 128K context; Mistral Research License (commercial use requires paid license); strong on multilingual + code generation." }, { "name": "o1 (preview & mini)", "family": "GPT / o-series", "org": "OpenAI", "release_date": "2024-09-12", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/learning-to-reason-with-llms/", "params": "unknown", "weights": "closed-api", "architecture": "decoder-only transformer (reinforcement-learning chain-of-thought)", "notes": "First major 'reasoning model'; spent inference compute on hidden chain-of-thought; PhD-level performance on physics/chem/bio; solved 83% of AIME problems vs GPT-4o's 13%." }, { "name": "Qwen2.5", "family": "Qwen", "org": "Alibaba Cloud", "release_date": "2024-09-19", "paper_arxiv_id": "2412.15115", "paper_url": "https://arxiv.org/abs/2412.15115", "params": "0.5B / 1.5B / 3B / 7B / 14B / 32B / 72B (+ Coder + Math + MoE)", "weights": "open", "architecture": "decoder-only transformer (+ MoE variants Turbo/Plus)", "notes": "Pretrained on 18T tokens (up from 7T); Qwen2.5-72B-Instruct rivaled Llama-3-405B at ~1/5 the parameters; paper landed Dec 19 2024." }, { "name": "Llama 3.2", "family": "LLaMA", "org": "Meta AI", "release_date": "2024-09-25", "paper_arxiv_id": null, "paper_url": "https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/", "params": "1B / 3B (text); 11B / 90B (vision)", "weights": "open", "architecture": "decoder-only transformer + vision adapters", "notes": "First multimodal Llama; 1B/3B for edge devices, 11B/90B with vision; not released in EU due to regulatory issues." }, { "name": "Claude 3.5 Sonnet (new) + Haiku + Computer Use", "family": "Claude", "org": "Anthropic", "release_date": "2024-10-22", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/3-5-models-and-computer-use", "params": "unknown", "weights": "closed-api", "architecture": "multimodal transformer", "notes": "Updated 3.5 Sonnet; first frontier model with 'Computer Use' beta (screen-perception + cursor + keyboard); SWE-bench Verified jumped from 33.4% to 49.0%." }, { "name": "Llama 3.3 70B", "family": "LLaMA", "org": "Meta AI", "release_date": "2024-12-06", "paper_arxiv_id": null, "paper_url": "https://ai.meta.com/blog/meta-llama-3-3/", "params": "70B", "weights": "open", "architecture": "decoder-only transformer", "notes": "Instruction-tuned-only release that matched Llama 3.1 405B at 4-5x lower serving cost; closing 2024's Llama cadence." }, { "name": "Gemini 2.0 Flash (Experimental)", "family": "Gemini", "org": "Google DeepMind", "release_date": "2024-12-11", "paper_arxiv_id": null, "paper_url": "https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/", "params": "unknown", "weights": "closed-api", "architecture": "multimodal transformer", "notes": "First Gemini 2 model; positioned for 'the agentic era'; native tool use; image and audio output baked in; GA on Feb 5 2025." }, { "name": "Phi-4", "family": "Phi", "org": "Microsoft Research", "release_date": "2024-12-12", "paper_arxiv_id": "2412.08905", "paper_url": "https://arxiv.org/abs/2412.08905", "params": "14B", "weights": "open", "architecture": "decoder-only transformer", "notes": "Synthetic-data-heavy training; beat GPT-4 on MATH and GPQA; weights MIT-licensed in early 2025." }, { "name": "DeepSeek-V3", "family": "DeepSeek", "org": "DeepSeek AI", "release_date": "2024-12-26", "paper_arxiv_id": "2412.19437", "paper_url": "https://arxiv.org/abs/2412.19437", "params": "671B total / 37B active", "weights": "open", "architecture": "MoE transformer (MLA + auxiliary-loss-free routing + multi-token prediction)", "notes": "Trained in only 2.788M H800 GPU-hours (~$5.6M compute); matched closed frontier models on many benchmarks; the model that triggered the 'DeepSeek moment' in January 2025." }, { "name": "DeepSeek-R1", "family": "DeepSeek", "org": "DeepSeek AI", "release_date": "2025-01-20", "paper_arxiv_id": "2501.12948", "paper_url": "https://arxiv.org/abs/2501.12948", "params": "671B total / 37B active (+ 1.5B-70B distilled variants)", "weights": "open", "architecture": "MoE transformer (pure-RL trained reasoning)", "notes": "First open reasoning model comparable to o1; demonstrated emergent CoT from RL alone (R1-Zero); the geopolitical shock-release that wiped >$1T from tech-stock market caps in late January 2025." }, { "name": "o3-mini", "family": "GPT / o-series", "org": "OpenAI", "release_date": "2025-01-31", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/openai-o3-mini/", "params": "unknown", "weights": "closed-api", "architecture": "decoder-only transformer (reasoning RL)", "notes": "Faster, cheaper reasoning model; available to free ChatGPT users (the first OpenAI reasoning model free-tier could access)." }, { "name": "Grok-3", "family": "Grok", "org": "xAI", "release_date": "2025-02-17", "paper_arxiv_id": null, "paper_url": "https://x.ai/news/grok-3", "params": "unknown (multi-trillion MoE rumored)", "weights": "closed-api", "architecture": "MoE transformer", "notes": "Trained on the Colossus supercluster (200K+ GPUs, 10x Grok-2 compute); introduced 'Think' and 'DeepSearch' reasoning modes." }, { "name": "Claude 3.7 Sonnet", "family": "Claude", "org": "Anthropic", "release_date": "2025-02-24", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-3-7-sonnet", "params": "unknown", "weights": "closed-api", "architecture": "multimodal transformer (hybrid reasoning)", "notes": "First hybrid reasoning model: same weights toggle between fast answers and extended thinking with a developer-controlled budget; first model to break 50% on SWE-bench Verified (62.3%)." }, { "name": "GPT-4.5 (Orion)", "family": "GPT", "org": "OpenAI", "release_date": "2025-02-27", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/introducing-gpt-4-5/", "params": "unknown (largest OpenAI model at release)", "weights": "closed-api", "architecture": "decoder-only transformer (dense)", "notes": "Internally code-named Orion; OpenAI's largest pretrained (non-reasoning) model; $75/$150 per million tokens \u2014 premium-priced even by frontier standards. Quietly deprecated mid-2025." }, { "name": "Gemini 2.5 Pro", "family": "Gemini", "org": "Google DeepMind", "release_date": "2025-03-25", "paper_arxiv_id": null, "paper_url": "https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/", "params": "unknown", "weights": "closed-api", "architecture": "multimodal transformer (thinking model)", "notes": "Google's first explicit 'thinking model'; chain-of-thought before responding; took #1 on LMSYS Chatbot Arena at launch." }, { "name": "Llama 4 (Scout + Maverick)", "family": "LLaMA", "org": "Meta AI", "release_date": "2025-04-05", "paper_arxiv_id": null, "paper_url": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/", "params": "Scout: 109B (17B active, 16 experts); Maverick: 400B (17B active, 128 experts)", "weights": "open", "architecture": "MoE multimodal transformer", "notes": "First Meta family to use MoE and natively multimodal; Scout has 10M-token context window (industry-leading at release); Behemoth (~2T params) remained in training." }, { "name": "o3 (full) + o4-mini", "family": "GPT / o-series", "org": "OpenAI", "release_date": "2025-04-16", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/introducing-o3-and-o4-mini/", "params": "unknown", "weights": "closed-api", "architecture": "decoder-only transformer (RL-trained reasoning + tool use)", "notes": "First OpenAI models with full agentic tool use in chain-of-thought (browser, Python, vision); o3 became the default 'paid frontier' reasoning model." }, { "name": "Qwen3", "family": "Qwen", "org": "Alibaba Cloud", "release_date": "2025-04-29", "paper_arxiv_id": "2505.09388", "paper_url": "https://arxiv.org/abs/2505.09388", "params": "0.6B / 1.7B / 4B / 8B / 14B / 32B (dense); 30B-A3B / 235B-A22B (MoE)", "weights": "open", "architecture": "decoder-only transformer (+ MoE variants; hybrid thinking)", "notes": "Hybrid thinking/non-thinking modes in a single model; trained on 36T tokens; the 235B-A22B MoE matched many closed frontier models on reasoning and coding." }, { "name": "Mistral Medium 3", "family": "Mistral", "org": "Mistral AI", "release_date": "2025-05-07", "paper_arxiv_id": null, "paper_url": "https://mistral.ai/news/mistral-medium-3", "params": "unknown", "weights": "closed-api", "architecture": "decoder-only transformer", "notes": "Targeted enterprise tier; 8x cheaper than comparable models at near-frontier quality; 128K context." }, { "name": "Claude Opus 4 + Sonnet 4", "family": "Claude", "org": "Anthropic", "release_date": "2025-05-22", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-4", "params": "unknown", "weights": "closed-api", "architecture": "multimodal hybrid-reasoning transformer", "notes": "First Claude family classified under ASL-3 safety measures; Opus 4 scored 72.5% on SWE-bench Verified; enabled background Claude Code runs." }, { "name": "DeepSeek-R1-0528", "family": "DeepSeek", "org": "DeepSeek AI", "release_date": "2025-05-28", "paper_arxiv_id": null, "paper_url": "https://api-docs.deepseek.com/news/news250528", "params": "671B total / 37B active", "weights": "open", "architecture": "MoE transformer (RL-trained reasoning)", "notes": "R1 minor-version refresh; AIME 2025 accuracy 70% -> 87.5%, hallucinations cut 45-50%; reasoning token budget effectively doubled." }, { "name": "Grok-4", "family": "Grok", "org": "xAI", "release_date": "2025-07-09", "paper_arxiv_id": null, "paper_url": "https://x.ai/news/grok-4", "params": "unknown", "weights": "closed-api", "architecture": "MoE transformer (multi-agent reasoning variant 'Heavy')", "notes": "Released after a livestreamed event; introduced 'Grok 4 Heavy' multi-agent reasoning variant; xAI's frontier push." }, { "name": "Kimi K2", "family": "Kimi", "org": "Moonshot AI", "release_date": "2025-07-11", "paper_arxiv_id": "2507.20534", "paper_url": "https://arxiv.org/abs/2507.20534", "params": "1T total / 32B active", "weights": "open", "architecture": "MoE transformer (Muon optimizer)", "notes": "First trillion-parameter open MoE optimized for agentic use; modified MIT license; trained with the Muon optimizer at unusual scale." }, { "name": "gpt-oss-120b / gpt-oss-20b", "family": "GPT", "org": "OpenAI", "release_date": "2025-08-05", "paper_arxiv_id": "2508.10925", "paper_url": "https://arxiv.org/abs/2508.10925", "params": "117B total / 5.1B active; 21B total / 3.6B active", "weights": "open", "architecture": "MoE transformer", "notes": "First OpenAI open-weight release since GPT-2 (6.5 years); Apache 2.0; 120B near-parity with o4-mini, runs on a single 80GB GPU." }, { "name": "GPT-5", "family": "GPT", "org": "OpenAI", "release_date": "2025-08-07", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/introducing-gpt-5/", "params": "unknown", "weights": "closed-api", "architecture": "router + multimodal transformer with reasoning", "notes": "Unified frontier model that auto-routes between 'thinking' and 'instant' modes; replaced the separate o-series + GPT-4o split in ChatGPT." }, { "name": "Claude Sonnet 4.5", "family": "Claude", "org": "Anthropic", "release_date": "2025-09-29", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-sonnet-4-5", "params": "unknown", "weights": "closed-api", "architecture": "multimodal hybrid-reasoning transformer", "notes": "Pitched as the best coding model in the world at release; capable of multi-hour autonomous coding sessions ('more of a colleague')." }, { "name": "Claude Haiku 4.5", "family": "Claude", "org": "Anthropic", "release_date": "2025-10-15", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-haiku-4-5", "params": "unknown", "weights": "closed-api", "architecture": "multimodal hybrid-reasoning transformer", "notes": "Hit Sonnet 4-level coding performance at 1/3 the price and >2x the speed; the small-tier in the Claude 4 generation." }, { "name": "Kimi K2 Thinking", "family": "Kimi", "org": "Moonshot AI", "release_date": "2025-11-06", "paper_arxiv_id": null, "paper_url": "https://moonshot.cn/blog/kimi-k2-thinking", "params": "1T total / 32B active", "weights": "open", "architecture": "MoE transformer (chain-of-thought reasoning)", "notes": "Open reasoning variant of Kimi K2; positioned by Moonshot as the most capable open thinking model at release." }, { "name": "GPT-5.1", "family": "GPT", "org": "OpenAI", "release_date": "2025-11-12", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/gpt-5-1/", "params": "unknown", "weights": "closed-api", "architecture": "router + multimodal transformer (adaptive reasoning)", "notes": "Introduced Instant + Thinking modes that adaptively allocate inference compute; 'warmer' default personality, customizable across 8 personas." }, { "name": "Gemini 3 Pro", "family": "Gemini", "org": "Google DeepMind", "release_date": "2025-11-18", "paper_arxiv_id": null, "paper_url": "https://blog.google/technology/google-deepmind/gemini-3/", "params": "unknown", "weights": "closed-api", "architecture": "multimodal transformer + Deep Think reasoning", "notes": "Top of LMSYS Arena at launch (1501 Elo); Deep Think mode hit 41% on Humanity's Last Exam; the model whose dominance triggered OpenAI's 'Code Red' that accelerated GPT-5.2." }, { "name": "Claude Opus 4.5", "family": "Claude", "org": "Anthropic", "release_date": "2025-11-24", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-opus-4-5", "params": "unknown", "weights": "closed-api", "architecture": "multimodal hybrid-reasoning transformer", "notes": "First model to break 80% on SWE-bench Verified (80.9%); price cut ~67% vs Opus 4.1 to $5/$25 per million tokens." }, { "name": "DeepSeek-V3.2", "family": "DeepSeek", "org": "DeepSeek AI", "release_date": "2025-12-01", "paper_arxiv_id": null, "paper_url": "https://api-docs.deepseek.com/news/news251201", "params": "~685B total / 37B active", "weights": "open", "architecture": "MoE transformer (Thinking-in-Tool-Use)", "notes": "Introduced native Thinking-in-Tool-Use; trained on a synthesized agentic-task pipeline covering 1800+ environments; ongoing R2 release still deferred per Liang." }, { "name": "Mistral Large 3", "family": "Mistral", "org": "Mistral AI", "release_date": "2025-12-02", "paper_arxiv_id": null, "paper_url": "https://mistral.ai/news/mistral-large-3", "params": "675B total / 41B active", "weights": "open", "architecture": "sparse MoE transformer", "notes": "Apache 2.0; Mistral's most capable open release; 256K context; competitive with closed frontier models." }, { "name": "GPT-5.2", "family": "GPT", "org": "OpenAI", "release_date": "2025-12-11", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/introducing-gpt-5-2/", "params": "unknown", "weights": "closed-api", "architecture": "router + multimodal transformer (adaptive reasoning)", "notes": "Released roughly three weeks after Gemini 3 Pro to reclaim frontier benchmarks; triggered by the internal 'Code Red' memo over Gemini 3's dominance." }, { "name": "Kimi K2.5", "family": "Kimi", "org": "Moonshot AI", "release_date": "2026-01-15", "paper_arxiv_id": null, "paper_url": "https://moonshot.cn/blog/kimi-k2-5", "params": "1T total / 32B active", "weights": "open", "architecture": "MoE multimodal transformer (agent swarm)", "notes": "Added native multimodal input and an agent-swarm paradigm; positioned by Moonshot as the open frontier of agentic AI." }, { "name": "Claude Opus 4.6", "family": "Claude", "org": "Anthropic", "release_date": "2026-02-04", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-opus-4-6", "params": "unknown", "weights": "closed-api", "architecture": "multimodal hybrid-reasoning transformer", "notes": "Incremental Opus refresh focused on agentic robustness and tool use; same $5/$25 pricing as 4.5." }, { "name": "GPT-5.3-Codex", "family": "GPT", "org": "OpenAI", "release_date": "2026-02-05", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/gpt-5-3-codex/", "params": "unknown", "weights": "closed-api", "architecture": "router + multimodal transformer (specialized coder)", "notes": "Coding-optimized 5.3 variant powering Codex Cloud and Codex CLI." }, { "name": "Claude Sonnet 4.6", "family": "Claude", "org": "Anthropic", "release_date": "2026-02-17", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-sonnet-4-6", "params": "unknown", "weights": "closed-api", "architecture": "multimodal hybrid-reasoning transformer", "notes": "Beta release shipped with a 1M-token context window \u2014 twice the previous Sonnet ceiling; aimed at Opus-level coding at Sonnet pricing ($3/$15)." }, { "name": "GPT-5.4", "family": "GPT", "org": "OpenAI", "release_date": "2026-03-05", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/gpt-5-4/", "params": "unknown", "weights": "closed-api", "architecture": "router + multimodal transformer", "notes": "Routine cadence frontier refresh; positioned between 5.2 and 5.5." }, { "name": "Claude Opus 4.7 (1M context)", "family": "Claude", "org": "Anthropic", "release_date": "2026-04-16", "paper_arxiv_id": null, "paper_url": "https://www.anthropic.com/news/claude-opus-4-7", "params": "unknown", "weights": "closed-api", "architecture": "multimodal hybrid-reasoning transformer", "notes": "1M-token context at standard pricing (no long-context premium); first Claude with high-resolution image support up to 2576px; 87.6% SWE-bench Verified, 94.2% GPQA." }, { "name": "Kimi K2.6", "family": "Kimi", "org": "Moonshot AI", "release_date": "2026-04-20", "paper_arxiv_id": null, "paper_url": "https://moonshot.cn/blog/kimi-k2-6", "params": "1T total / 32B active", "weights": "open", "architecture": "MoE multimodal transformer (long-horizon agent swarm)", "notes": "Agent swarm scaled to 300 sub-agents and 4000 coordinated steps; long-horizon coding focus." }, { "name": "GPT-5.5", "family": "GPT", "org": "OpenAI", "release_date": "2026-04-23", "paper_arxiv_id": null, "paper_url": "https://openai.com/index/introducing-gpt-5-5/", "params": "unknown", "weights": "closed-api", "architecture": "router + multimodal transformer", "notes": "Bridge toward an OpenAI 'super app'; Instant variant became the default ChatGPT model on May 5." }, { "name": "Gemini 3.5 Flash", "family": "Gemini", "org": "Google DeepMind", "release_date": "2026-05-19", "paper_arxiv_id": null, "paper_url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/", "params": "unknown", "weights": "closed-api", "architecture": "multimodal transformer (agentic specialization)", "notes": "Launched at Google I/O 2026; outperforms Gemini 3.1 Pro on agentic + multimodal benchmarks at ~4x faster inference and <50% the cost ($1.50/$9.00 per million tokens)." } ]