{
 "models": [
  {
   "name": "RoBERTa-large",
   "project_id": "roberta",
   "date": "2019-07",
   "params_b": 0.355,
   "modality": "text",
   "open": true,
   "note": "BERT-large architecture retrained on 160GB; base is 125M",
   "source": "https://github.com/facebookresearch/fairseq/tree/main/examples/roberta"
  },
  {
   "name": "BART-large",
   "project_id": "bart",
   "date": "2019-10",
   "params_b": 0.4,
   "modality": "text",
   "open": true,
   "note": "12-layer denoising encoder-decoder; base is 140M",
   "source": "https://github.com/facebookresearch/fairseq/tree/main/examples/bart"
  },
  {
   "name": "wav2vec 2.0 LARGE",
   "project_id": "wav2vec2",
   "date": "2020-06",
   "params_b": 0.317,
   "modality": "speech",
   "open": true,
   "note": "LARGE 317M, BASE 95M; self-supervised speech pretraining",
   "source": "https://arxiv.org/abs/2006.11477"
  },
  {
   "name": "XLM-R XXL",
   "project_id": "xlmr",
   "date": "2021-05",
   "params_b": 10.7,
   "modality": "text",
   "open": true,
   "note": "48-layer scale-up of 2019's XLM-R (Large 550M, XL 3.5B); 100 languages",
   "source": "https://arxiv.org/abs/2105.00572"
  },
  {
   "name": "OPT-125M",
   "project_id": "opt",
   "date": "2022-05",
   "params_b": 0.125,
   "modality": "text",
   "open": true,
   "note": "Smallest of the OPT suite (125M to 175B)",
   "source": "https://arxiv.org/abs/2205.01068"
  },
  {
   "name": "OPT-175B",
   "project_id": "opt",
   "date": "2022-05",
   "params_b": 175,
   "modality": "text",
   "open": true,
   "note": "GPT-3 scale; 125M-66B freely released, 175B by research request; training logbook published",
   "source": "https://arxiv.org/abs/2205.01068"
  },
  {
   "name": "NLLB-200 (MoE)",
   "project_id": "nllb",
   "date": "2022-07",
   "params_b": 54.5,
   "modality": "text",
   "open": true,
   "note": "Sparsely gated Mixture-of-Experts with FLOPs of a 3.3B dense model; 200 languages",
   "source": "https://arxiv.org/abs/2207.04672",
   "active_b": 3.3
  },
  {
   "name": "NLLB-200 3.3B",
   "project_id": "nllb",
   "date": "2022-07",
   "params_b": 3.3,
   "modality": "text",
   "open": true,
   "note": "Dense NLLB-200 variant (also 1.3B dense)",
   "source": "https://huggingface.co/facebook/nllb-200-3.3B"
  },
  {
   "name": "NLLB-200 distilled 600M",
   "project_id": "nllb",
   "date": "2022-07",
   "params_b": 0.6,
   "modality": "text",
   "open": true,
   "note": "Smallest distilled NLLB-200 checkpoint",
   "source": "https://huggingface.co/facebook/nllb-200-distilled-600M"
  },
  {
   "name": "BlenderBot 3 175B",
   "project_id": "blenderbot",
   "date": "2022-08",
   "params_b": 175,
   "modality": "text",
   "open": true,
   "note": "Initialized from OPT-175B; also 3B and 30B; public US demo months before ChatGPT",
   "source": "https://arxiv.org/abs/2208.03188"
  },
  {
   "name": "ESM-2 15B",
   "project_id": "esm",
   "date": "2022-08",
   "params_b": 15,
   "modality": "science",
   "open": true,
   "note": "Largest ESM-2 protein language model (esm2_t48_15B_UR50D); family starts at 8M",
   "source": "https://github.com/facebookresearch/esm"
  },
  {
   "name": "ESMFold",
   "project_id": "esm",
   "date": "2022-11",
   "params_b": 3.7,
   "modality": "science",
   "open": true,
   "note": "690M folding trunk on a 3B ESM-2 backbone (esmfold_v1); folded 617M metagenomic proteins",
   "source": "https://github.com/facebookresearch/esm"
  },
  {
   "name": "Galactica 120B",
   "project_id": "galactica",
   "date": "2022-11",
   "params_b": 120,
   "modality": "text",
   "open": true,
   "note": "Science LLM, sizes 125M-120B; demo pulled after 3 days, weights stayed up",
   "source": "https://arxiv.org/abs/2211.09085"
  },
  {
   "name": "CICERO",
   "project_id": "cicero",
   "date": "2022-11",
   "params_b": 2.7,
   "modality": "text",
   "open": true,
   "note": "2.7B BART-like dialogue model paired with a strategic planner; human-level Diplomacy",
   "source": "https://ai.meta.com/blog/cicero-ai-negotiates-persuades-and-cooperates-with-people/"
  },
  {
   "name": "LLaMA 7B",
   "project_id": "llama",
   "date": "2023-02",
   "params_b": 6.7,
   "modality": "text",
   "open": true,
   "note": "Research-only weights (gated), leaked within a week",
   "source": "https://arxiv.org/abs/2302.13971"
  },
  {
   "name": "LLaMA 13B",
   "project_id": "llama",
   "date": "2023-02",
   "params_b": 13.0,
   "modality": "text",
   "open": true,
   "note": "Research-only weights (gated), leaked within a week; 13B beat GPT-3 175B on many benchmarks",
   "source": "https://arxiv.org/abs/2302.13971"
  },
  {
   "name": "LLaMA 33B",
   "project_id": "llama",
   "date": "2023-02",
   "params_b": 32.5,
   "modality": "text",
   "open": true,
   "note": "Research-only weights (gated), leaked within a week",
   "source": "https://arxiv.org/abs/2302.13971"
  },
  {
   "name": "LLaMA 65B",
   "project_id": "llama",
   "date": "2023-02",
   "params_b": 65.2,
   "modality": "text",
   "open": true,
   "note": "Research-only weights (gated), leaked within a week; trained on 1.4T tokens",
   "source": "https://arxiv.org/abs/2302.13971"
  },
  {
   "name": "SAM (ViT-H)",
   "project_id": "sam",
   "date": "2023-04",
   "params_b": 0.636,
   "modality": "vision",
   "open": true,
   "note": "ViT-H image encoder 636M (ViT-L 308M, ViT-B 91M); Apache 2.0",
   "source": "https://arxiv.org/abs/2304.02643"
  },
  {
   "name": "DINOv2 ViT-g/14",
   "project_id": "dinov2",
   "date": "2023-04",
   "params_b": 1.1,
   "modality": "vision",
   "open": true,
   "note": "Self-supervised backbone; distilled ViT-S/B/L at 21M/86M/300M",
   "source": "https://github.com/facebookresearch/dinov2"
  },
  {
   "name": "ImageBind-huge",
   "project_id": "imagebind",
   "date": "2023-05",
   "params_b": 0.63,
   "modality": "multimodal",
   "open": true,
   "note": "Documented figure is the ViT-H image encoder (630M) plus a 302M text encoder from OpenCLIP; Meta never published a total for the six-modality checkpoint",
   "source": "https://arxiv.org/abs/2305.05665"
  },
  {
   "name": "MMS-1B",
   "project_id": "mms",
   "date": "2023-05",
   "params_b": 1.0,
   "modality": "speech",
   "open": true,
   "note": "wav2vec 2.0 scaled to 1B; ASR for 1,100+ languages",
   "source": "https://huggingface.co/facebook/mms-1b-all"
  },
  {
   "name": "I-JEPA ViT-H",
   "project_id": "ijepa",
   "date": "2023-06",
   "params_b": 0.632,
   "modality": "vision",
   "open": true,
   "note": "Trained in under 72 hours on 16 A100s; first JEPA release",
   "source": "https://ai.meta.com/blog/yann-lecun-ai-model-i-jepa/"
  },
  {
   "name": "MusicGen small",
   "project_id": "audiocraft",
   "date": "2023-06",
   "params_b": 0.3,
   "modality": "audio",
   "open": true,
   "note": "Text-to-music; weights CC-BY-NC",
   "source": "https://github.com/facebookresearch/audiocraft/blob/main/docs/MUSICGEN.md"
  },
  {
   "name": "MusicGen medium",
   "project_id": "audiocraft",
   "date": "2023-06",
   "params_b": 1.5,
   "modality": "audio",
   "open": true,
   "note": "Text-to-music; weights CC-BY-NC",
   "source": "https://github.com/facebookresearch/audiocraft/blob/main/docs/MUSICGEN.md"
  },
  {
   "name": "MusicGen large",
   "project_id": "audiocraft",
   "date": "2023-06",
   "params_b": 3.3,
   "modality": "audio",
   "open": true,
   "note": "Largest MusicGen; weights CC-BY-NC",
   "source": "https://github.com/facebookresearch/audiocraft/blob/main/docs/MUSICGEN.md"
  },
  {
   "name": "Voicebox",
   "project_id": "voicebox",
   "date": "2023-06",
   "params_b": 0.33,
   "modality": "speech",
   "open": false,
   "note": "Never released; Meta withheld weights over voice-cloning risk",
   "source": "https://arxiv.org/abs/2306.15687"
  },
  {
   "name": "Llama 2 7B",
   "project_id": "llama2",
   "date": "2023-07",
   "params_b": 7.0,
   "modality": "text",
   "open": true,
   "note": "First commercially licensed Llama; a 34B variant was trained but never released",
   "source": "https://arxiv.org/abs/2307.09288"
  },
  {
   "name": "Llama 2 13B",
   "project_id": "llama2",
   "date": "2023-07",
   "params_b": 13.0,
   "modality": "text",
   "open": true,
   "note": "First commercially licensed Llama; a 34B variant was trained but never released",
   "source": "https://arxiv.org/abs/2307.09288"
  },
  {
   "name": "Llama 2 70B",
   "project_id": "llama2",
   "date": "2023-07",
   "params_b": 70.0,
   "modality": "text",
   "open": true,
   "note": "First commercially licensed Llama; a 34B variant was trained but never released",
   "source": "https://arxiv.org/abs/2307.09288"
  },
  {
   "name": "Code Llama 7B",
   "project_id": "code-llama",
   "date": "2023-08",
   "params_b": 7.0,
   "modality": "text",
   "open": true,
   "note": "Llama 2 specialized for code; base, Python and Instruct variants",
   "source": "https://arxiv.org/abs/2308.12950"
  },
  {
   "name": "Code Llama 13B",
   "project_id": "code-llama",
   "date": "2023-08",
   "params_b": 13.0,
   "modality": "text",
   "open": true,
   "note": "Llama 2 specialized for code; base, Python and Instruct variants",
   "source": "https://arxiv.org/abs/2308.12950"
  },
  {
   "name": "Code Llama 34B",
   "project_id": "code-llama",
   "date": "2023-08",
   "params_b": 34.0,
   "modality": "text",
   "open": true,
   "note": "Llama 2 specialized for code; base, Python and Instruct variants",
   "source": "https://arxiv.org/abs/2308.12950"
  },
  {
   "name": "Emu",
   "project_id": "emu",
   "date": "2023-09",
   "params_b": 2.8,
   "modality": "vision",
   "open": false,
   "note": "2.8B U-Net latent diffusion image generator; closed product model behind Meta AI 'Imagine'",
   "source": "https://arxiv.org/abs/2309.15807"
  },
  {
   "name": "SeamlessM4T v2 Large",
   "project_id": "seamless",
   "date": "2023-11",
   "params_b": 2.3,
   "modality": "speech",
   "open": true,
   "note": "Speech/text any-to-any translation across ~100 languages; v1 Large (Aug 2023) was also 2.3B",
   "source": "https://huggingface.co/facebook/seamless-m4t-v2-large"
  },
  {
   "name": "Code Llama 70B",
   "project_id": "code-llama",
   "date": "2024-01",
   "params_b": 70,
   "modality": "text",
   "open": true,
   "note": "Added Jan 29, 2024; best open code model at release (67.8% HumanEval)",
   "source": "https://ai.meta.com/blog/code-llama-large-language-model-coding/"
  },
  {
   "name": "MobileLLM-125M",
   "project_id": "mobilellm",
   "date": "2024-02",
   "params_b": 0.125,
   "modality": "text",
   "open": true,
   "note": "Deep-and-thin sub-billion design; weights public on HF Oct 30, 2024 (CC-BY-NC)",
   "source": "https://github.com/facebookresearch/MobileLLM"
  },
  {
   "name": "MobileLLM-1.5B",
   "project_id": "mobilellm",
   "date": "2024-02",
   "params_b": 1.5,
   "modality": "text",
   "open": true,
   "note": "Largest MobileLLM; weights public on HF Oct 30, 2024 (CC-BY-NC)",
   "source": "https://github.com/facebookresearch/MobileLLM"
  },
  {
   "name": "Llama 3 8B",
   "project_id": "llama3",
   "date": "2024-04",
   "params_b": 8,
   "modality": "text",
   "open": true,
   "note": "Trained on 15T tokens; 128K-vocab tokenizer",
   "source": "https://ai.meta.com/blog/meta-llama-3/"
  },
  {
   "name": "Llama 3 70B",
   "project_id": "llama3",
   "date": "2024-04",
   "params_b": 70,
   "modality": "text",
   "open": true,
   "note": "Trained on two 24,576-GPU clusters",
   "source": "https://ai.meta.com/blog/meta-llama-3/"
  },
  {
   "name": "Chameleon 7B",
   "project_id": "chameleon",
   "date": "2024-05",
   "params_b": 7,
   "modality": "multimodal",
   "open": true,
   "note": "Early-fusion mixed-modal transformer; Chameleon Research License, image output disabled",
   "source": "https://arxiv.org/abs/2405.09818"
  },
  {
   "name": "Chameleon 34B",
   "project_id": "chameleon",
   "date": "2024-05",
   "params_b": 34,
   "modality": "multimodal",
   "open": true,
   "note": "Released on Hugging Face as 'chameleon-30b' (34B params); Chameleon Research License",
   "source": "https://huggingface.co/facebook/chameleon-30b"
  },
  {
   "name": "Llama 3.1 405B",
   "project_id": "llama3",
   "date": "2024-07",
   "params_b": 405,
   "modality": "text",
   "open": true,
   "note": "First open frontier-class model; >15T tokens on 16,000+ H100s",
   "source": "https://ai.meta.com/blog/meta-llama-3-1/"
  },
  {
   "name": "SAM 2 (Hiera-L)",
   "project_id": "sam2",
   "date": "2024-07",
   "params_b": 0.2244,
   "modality": "vision",
   "open": true,
   "note": "sam2_hiera_large 224.4M; SAM 2.1 refresh (Sept 2024) kept the same sizes; tiny is 38.9M",
   "source": "https://github.com/facebookresearch/sam2"
  },
  {
   "name": "Llama 3.2 1B",
   "project_id": "llama3",
   "date": "2024-09",
   "params_b": 1,
   "modality": "text",
   "open": true,
   "note": "Edge/mobile text model",
   "source": "https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/"
  },
  {
   "name": "Llama 3.2 3B",
   "project_id": "llama3",
   "date": "2024-09",
   "params_b": 3,
   "modality": "text",
   "open": true,
   "note": "Edge/mobile text model",
   "source": "https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/"
  },
  {
   "name": "Llama 3.2 11B Vision",
   "project_id": "llama3",
   "date": "2024-09",
   "params_b": 11,
   "modality": "multimodal",
   "open": true,
   "note": "First Llama with image input",
   "source": "https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/"
  },
  {
   "name": "Llama 3.2 90B Vision",
   "project_id": "llama3",
   "date": "2024-09",
   "params_b": 90,
   "modality": "multimodal",
   "open": true,
   "note": "Largest Llama vision model",
   "source": "https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/"
  },
  {
   "name": "Movie Gen Video",
   "project_id": "movie-gen",
   "date": "2024-10",
   "params_b": 30,
   "modality": "video",
   "open": false,
   "note": "30B transformer, 16s of 1080p; weights never released",
   "source": "https://arxiv.org/abs/2410.13720"
  },
  {
   "name": "Movie Gen Audio",
   "project_id": "movie-gen",
   "date": "2024-10",
   "params_b": 13,
   "modality": "audio",
   "open": false,
   "note": "13B video-to-audio DiT; weights never released",
   "source": "https://arxiv.org/abs/2410.13720"
  },
  {
   "name": "Spirit LM",
   "project_id": "spirit-lm",
   "date": "2024-10",
   "params_b": 7,
   "modality": "multimodal",
   "open": true,
   "note": "Llama-2 7B continued on interleaved text+speech tokens; paper Feb 2024, weights Oct 2024 (FAIR NC)",
   "source": "https://arxiv.org/abs/2402.05755"
  },
  {
   "name": "Llama 3.3 70B",
   "project_id": "llama3",
   "date": "2024-12",
   "params_b": 70,
   "modality": "text",
   "open": true,
   "note": "405B-class quality at 70B; instruct-only release Dec 6, 2024",
   "source": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct"
  },
  {
   "name": "Large Concept Model 7B",
   "project_id": "lcm",
   "date": "2024-12",
   "params_b": 7,
   "modality": "text",
   "open": false,
   "note": "Sentence-level 'concept' prediction in SONAR space; training code open, no pretrained weights (also 1.6B)",
   "source": "https://arxiv.org/abs/2412.08821"
  },
  {
   "name": "Byte Latent Transformer 8B",
   "project_id": "blt",
   "date": "2024-12",
   "params_b": 8,
   "modality": "text",
   "open": true,
   "note": "Tokenizer-free byte model; paper scaled to 8B, gated HF weights shipped as blt-1b and blt-7b",
   "source": "https://arxiv.org/abs/2412.09871"
  },
  {
   "name": "Llama 4 Scout",
   "project_id": "llama4",
   "date": "2025-04",
   "params_b": 109,
   "modality": "multimodal",
   "open": true,
   "note": "16-expert MoE, 17B active; 10M-token context",
   "source": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/",
   "active_b": 17
  },
  {
   "name": "Llama 4 Maverick",
   "project_id": "llama4",
   "date": "2025-04",
   "params_b": 400,
   "modality": "multimodal",
   "open": true,
   "note": "128-expert MoE, 17B active",
   "source": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/",
   "active_b": 17
  },
  {
   "name": "Llama 4 Behemoth",
   "project_id": "llama4",
   "date": "2025-04",
   "params_b": 2000,
   "modality": "multimodal",
   "open": false,
   "note": "Never released; 'nearly two trillion' total, 288B active, 16 experts; used as a teacher",
   "source": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/",
   "active_b": 288
  },
  {
   "name": "Perception Encoder Core G/14",
   "project_id": "perception",
   "date": "2025-04",
   "params_b": 1.88,
   "modality": "vision",
   "open": true,
   "note": "Vision tower 1.88B (+0.47B text tower); CLIP-style encoder, best features in intermediate layers",
   "source": "https://huggingface.co/facebook/PE-Core-G14-448"
  },
  {
   "name": "V-JEPA 2 ViT-g",
   "project_id": "vjepa2",
   "date": "2025-06",
   "params_b": 1,
   "modality": "world",
   "open": true,
   "note": "Video world model trained on 1M+ hours; ViT-L 300M and ViT-H 600M also released",
   "source": "https://github.com/facebookresearch/vjepa2"
  },
  {
   "name": "UMA-M",
   "project_id": "uma",
   "date": "2025-06",
   "params_b": 1.4,
   "modality": "science",
   "open": true,
   "note": "Mixture of Linear Experts interatomic potential, ~50M active per structure; UMA-S is 150M/6M active",
   "source": "https://arxiv.org/abs/2506.23971",
   "active_b": 0.05
  },
  {
   "name": "DINOv3 ViT-7B/16",
   "project_id": "dinov3",
   "date": "2025-08",
   "params_b": 6.7,
   "modality": "vision",
   "open": true,
   "note": "6,716M self-supervised backbone trained on 1.7B images; distilled ViT-H+ 840M, ViT-L 300M",
   "source": "https://github.com/facebookresearch/dinov3"
  },
  {
   "name": "MobileLLM-R1-950M",
   "project_id": "mobilellm-r1",
   "date": "2025-09",
   "params_b": 0.949,
   "modality": "text",
   "open": true,
   "note": "Sub-billion reasoning model, full open recipe; FAIR NC license",
   "source": "https://huggingface.co/facebook/MobileLLM-R1-950M"
  },
  {
   "name": "Code World Model",
   "project_id": "cwm",
   "date": "2025-09",
   "params_b": 32,
   "modality": "text",
   "open": true,
   "note": "32B dense research model of code execution; FAIR NC license, 3 training-stage checkpoints",
   "source": "https://huggingface.co/facebook/cwm"
  },
  {
   "name": "MobileLLM-Pro",
   "project_id": "mobilellm-pro",
   "date": "2025-10",
   "params_b": 1.08,
   "modality": "text",
   "open": true,
   "note": "Reality Labs 1.08B on-device model, 128k context; FAIR NC license",
   "source": "https://huggingface.co/facebook/MobileLLM-Pro"
  },
  {
   "name": "Omnilingual ASR LLM-7B",
   "project_id": "omnilingual-asr",
   "date": "2025-11",
   "params_b": 7.8,
   "modality": "speech",
   "open": true,
   "note": "omniASR_LLM_7B is 7.80B (its 7B w2v encoder alone is 6.49B); suite spans 300M-7B, 1,600+ languages, Apache 2.0",
   "source": "https://github.com/facebookresearch/omnilingual-asr"
  },
  {
   "name": "SAM 3",
   "project_id": "sam3",
   "date": "2025-11",
   "params_b": 0.848,
   "modality": "vision",
   "open": true,
   "note": "Detector + tracker sharing one vision encoder; text-promptable segmentation",
   "source": "https://github.com/facebookresearch/sam3"
  },
  {
   "name": "V-JEPA 2.1 ViT-G",
   "project_id": "vjepa2",
   "date": "2026-03",
   "params_b": 2,
   "modality": "world",
   "open": true,
   "note": "V-JEPA 2.1 family spans ViT-B 80M to ViT-G 2B",
   "source": "https://github.com/facebookresearch/vjepa2"
  },
  {
   "name": "Muse Glimmer 30B",
   "project_id": "muse-glimmer",
   "date": "2026-08",
   "params_b": 29.6,
   "modality": "multimodal",
   "open": true,
   "note": "Dense 29.6B (incl. vision encoder) distilled from Muse Spark; Apache 2.0; Meta's return to open weights. Muse Spark itself is closed with no published parameter count",
   "source": "https://huggingface.co/meta-models/Muse-Glimmer-30B"
  }
 ]
}