The Meta AI Universe

All 119 projects

The star map above needs JavaScript. This is the same catalog as a plain list, grouped the way the sky groups it. Each name links to that project's own page, which carries its sources, its lineage and its live figures. Take the whole thing as a CSV, or read the timeline, the comparison and the method, none of which need scripting either.

Language & LLMs 20

• Llama 4 2025 open release · superseded
  A 10-million-token context window — and a turning point
• Code World Model 2025 open release
  A neural debugger that simulates code in its head
• Llama API 2025 closed · archived
  Meta's hosted Llama, folded into the Model API in 2026
• Llama 3 / 3.1 / 3.2 / 3.3 2024 open release · superseded
  405B: the first GPT-4-class model anyone could download
• Large Concept Models 2024 open release
  An AI that predicts the next idea, not the next word
• Byte Latent Transformer 2024 open release
  The architecture that killed the tokenizer
• Llama Stack 2024 open release
  Write the app once, run it local, cloud or on-prem
• Coconut 2024 open release
  Reasoning in latent space instead of words
• Multi-token prediction 2024 open release
  Several tokens per step: stronger code models, up to 3x faster
• Memory Layers at Scale 2024 open release
  Capacity without FLOPs: a lookup table the model learns to consult
• LLaMA 2023 open release · superseded
  The leak that started the open-weight revolution
• Llama 2 2023 open release · superseded
  Open weights you could build a business on
• Code Llama 2023 open release · superseded
  Llama learns to program
• OPT-175B 2022 open release · superseded
  The 175B model that published its 3 a.m. crash logs
• Galactica 2022 open release · archived
  The science LLM that lasted three days in public
• BlenderBot 2020 open release · archived
  A public chatbot that criticized its own CEO — months before ChatGPT
• RoBERTa 2019 open release · superseded
  BERT, done right
• XLM-R 2019 open release · superseded
  100 languages in one encoder — the ancestor of NLLB
• BART 2019 open release · superseded
  Corrupt the text, learn to rebuild it — summarization's workhorse
• fairseq 2017 open release · archived
  The toolkit that trained a decade of breakthroughs

Vision 17

• SAM 3 2025 open release
  Type 'yellow school bus' — it masks every single one
• SAM 3D (Objects + Body) 2025 open release
  A single photo becomes a 3D object
• DINOv3 2025 open release
  One frozen backbone, from Instagram photos to Mars
• VGGT 2025 open release
  CVPR 2025 Best Paper: geometry without the geometry pipeline
• Perception Encoder & Perception Language Model 2025 open release
  A vision encoder whose best features hide in its middle layers
• SAM 2 2024 open release · superseded
  One click on one frame tracks an object through a whole video
• Chameleon 2024 open release
  Images and text in one token stream, trained from scratch
• Sapiens 2024 open release
  308 keypoints per body, learned from 300 million human images
• Video Seal 2024 open release
  Invisible video watermarks that survive blur, crops and compression
• Segment Anything 2023 open release · superseded
  Click anything, cut out anything — segmentation, solved
• DINOv2 2023 open release · superseded
  Label-free features that beat supervised pipelines
• ImageBind 2023 open release
  One embedding space to bind six senses
• CoTracker 2023 open release
  Follow any pixel through a video — even when it disappears
• DINO 2021 open release · superseded
  Attention maps that learned to see objects — with zero labels
• PyTorch3D 2020 open release
  Fit a 3D mesh from photos by gradient descent
• Detectron2 2019 open release
  Detection and segmentation, boxed up as the field's default toolkit
• Mask R-CNN 2017 open release · superseded
  The detection workhorse of an era

World Models & Embodied AI 12

• V-JEPA 2 2025 open release
  It learned physics from video — then drove a robot arm it had never met
• VL-JEPA 2025 closed
  Predict the embedding, skip the tokens: JEPA outlives its champion
• Meta Locate 3D 2025 open release
  'The flower vase near the TV console' — located in 3D
• V-JEPA 2024 open release · superseded
  Released the same month as Sora — and betting the opposite way
• PARTNR 2024 open release
  100,000 household chores that showed robot planners slow humans down
• OpenEQA 2024 open release
  1,600 questions about real spaces — top VLMs were 'nearly blind'
• Meta Motivo 2024 open release
  Puppet a physics-based humanoid with a prompt
• I-JEPA 2023 open release · superseded
  LeCun's world-model bet, realized: predict representations, not pixels
• Ego-Exo4D 2023 open release
  Cooking, bike repair, bouldering — seen first- and third-person at once
• Ego4D 2021 open release
  3,670 hours of life, seen through human eyes
• Project Aria 2020 closed
  Research glasses for teaching machines to see like us
• Habitat 2019 open release
  A simulator where robots learn to walk before they exist

Speech & Sound 12

• Omnilingual ASR 2025 open release
  Speech recognition for 1,600+ languages — most for the first time
• SAM Audio 2025 open release
  Segment Anything — for sound
• Language Technology Partner Program + BOUQuET 2025 open release
  Partners contribute underserved languages; the resulting models go open
• Spirit LM 2024 open release
  Text and speech in one token stream — with the emotion intact
• SeamlessM4T & the Seamless family 2023 open release
  A real-life Babel Fish that keeps your voice
• Massively Multilingual Speech 2023 open release · superseded
  Speech tech for 1,100+ languages
• AudioCraft 2023 open release
  Type a vibe, get a song
• Voicebox 2023 closed
  So good at voices Meta wouldn't release it
• Audiobox 2023 closed
  Describe the voice and the room — it generates both
• No Language Left Behind 2022 open release
  200 languages, one model, published in Nature
• Universal Speech Translator 2022 open release · superseded
  Translating a language with no standard written form
• wav2vec 2.0 2020 open release · superseded
  Speech recognition from raw audio, almost no labels

Open Source Infra 11

• ExecuTorch 2023 open release
  PyTorch in your pocket
• Animated Drawings 2021 open release
  Your kid's doodle, walking
• xFormers 2021 open release
  Exact attention without the O(n²) memory bill
• Hydra 2019 open release
  Compose YAML configs, override from the shell, sweep in one flag
• Nevergrad 2018 open release
  Dozens of black-box optimizers behind one ask/tell API
• faiss 2017 open release
  The quiet backbone of the vector-database era
• Prophet 2017 open release
  The forecasting tool every analyst has met
• PyTorch Domain Libraries 2017 open release
  torchvision, torchaudio, torchtune — the batteries that came with PyTorch
• ParlAI 2017 open release · archived
  The dialogue framework that raised BlenderBot, archived in the LLM era
• PyTorch 2016 open release
  The framework the AI world runs on
• fastText 2016 open release · archived
  Language detection for 176 languages in under a megabyte

The Muse Era 9

• Muse Spark 2026 closed
  Meta's first closed frontier model — and its comeback story
• Muse Glimmer 2026 open release
  The open-source plot twist: a 30B agent that runs on your laptop
• Muse Image 2026 closed
  An image model that thinks before it draws
• Muse Video 2026 closed · preview
  The Movie Gen lineage goes to production
• Muse Code 2026 closed
  Meta's terminal coding agent, powered by Spark 1.2
• Meta Model API 2026 closed
  Muse Spark for developers
• Meta Superintelligence Labs 2025 closed
  The 2025 reboot: a superintelligence lab built around a $14.3B talent bet
• Meta AI app 2025 closed
  The assistant in front of 3 billion people
• Vibes 2025 closed
  A TikTok-style feed where every video is AI-made

AI for Science 9

• Open Molecules 2025 2025 open release
  100 million quantum-chemistry calculations, given away
• UMA — Universal Models for Atoms 2025 open release
  ~10,000x faster than DFT — a universal model for atoms
• Brain2Qwerty 2025 open release
  Typing decoded from non-invasive brain recordings
• Open Materials 2024 2024 open release
  Open data that topped Matbench Discovery for predicting stable crystals
• fairchem 2024 open release
  One pip install for five generations of FAIR chemistry
• OpenDAC / ODAC25 2023 open release
  70 million quantum calculations for materials that pull CO₂ from air
• Brain & AI: Decoding Perception from MEG/EEG 2022 open release
  Reconstructing what you see from brain signals
• Open Catalyst Project 2020 open release
  Hunting clean-energy catalysts with AI
• ESM / ESMFold 2019 open release · archived
  617 million protein structures in two weeks

Reality Labs Research 9

• Generic Neuromotor Interface 2025 open release
  A wristband that reads intention from your nerves
• Meta Ray-Ban Display + Neural Band 2025 closed
  The neural wristband ships
• Meta AI Glasses lineup 2025 closed
  Camera-and-voice glasses from Ray-Ban and Oakley, now running Muse Spark
• Orion 2024 closed
  The AR glasses prototype a decade in the making
• Codec Avatar Studio 2024 open release
  Paired dome scans and headset footage of 256 people, released open
• audio2photoreal 2024 open release
  Record your voice, watch a photoreal avatar gesture along
• emg2qwerty 2024 open release
  Every keystroke of 108 typists, paired with their wrist signals
• emg2pose 2024 open release
  Hand tracking from the wrist — no camera required
• Codec Avatars 2019 closed
  Telepresence that looks exactly like you

On-Device & Silicon 8

• MobileLLM-Flash 2026 closed
  Architecture search with real phone latency in the loop
• MobileLLM-R1 / R1.5 2025 open release
  Sub-billion reasoners matching Qwen3-0.6B on a seventh of the tokens
• MobileLLM-Pro 2025 open release
  1B parameters, 128k context — built by the smart-glasses org
• MobileLLM 2024 open release
  Proof that sub-billion models could think
• GPU superclusters: 350K H100s → Prometheus & Hyperion 2024 closed
  Gigawatt-scale factories for intelligence
• Quantized Llama 3.2 2024 open release
  Llama 3.2 on a phone — 56% smaller, up to 4x faster
• Meta LLM Compiler 2024 open release
  77% of a full autotuning search's gains, without compiling once
• MTIA 2023 closed
  Meta's own AI silicon

Generative Media 4

• Movie Gen 2024 closed · superseded
  A sentence in, a 16-second film with its own score out
• Emu 2023 closed · superseded
  A few thousand perfect photos beat millions of mediocre ones
• Emu Video & Emu Edit 2023 closed · superseded
  Two Emu offshoots: 4-second clips and edit-by-instruction — demos only
• Make-A-Video 2022 closed · superseded
  Text-to-video before it was normal

Games & Strategy 4

• CICERO 2022 open release
  It won Diplomacy by out-talking humans — and no one called it out
• ReBeL 2020 open release
  AlphaZero for hidden-information games — the poker code stayed locked
• Pluribus 2019 closed
  Superhuman six-player poker for $150 of compute
• ELF OpenGo 2018 open release · archived
  20-0 against Go professionals, on a single GPU — then open-sourced

Safety & Trust 4

• LlamaFirewall 2025 open release
  Prompt-injection success rate: 17.6% without, 1.7% with
• Purple Llama 2023 open release
  Open-source safety tooling for open models
• CyberSecEval 2023 open release
  Can your model break code — and can it patch it?
• GAIA Benchmark 2023 open release
  466 questions: humans scored 92%, GPT-4 with plugins scored 15%