Thirteen years, six acts
From a research lab wagering on deep learning to the Muse era — the whole documented arc, with the drama left in. Titles link to stars in the Universe.
First Light
In December 2013 Mark Zuckerberg hires NYU professor Yann LeCun — one of the godfathers of deep learning — to found FAIR, Facebook AI Research, with an unusual promise: the research would be open. The bet pays off almost immediately, in tools the whole field ends up using.
FAIR is founded
Yann LeCun opens Facebook AI Research with labs in New York, Menlo Park and London, committed to open publication — a culture that will define everything that follows.
fastText
Word vectors and text classification fast enough to run on a laptop — eventually spanning 157 languages. For years, the default first tool of applied NLP.
PyTorch is born
A small FAIR team rethinks what a deep-learning framework should feel like: define-by-run, Pythonic, debuggable. Within five years it becomes the language of AI research itself.
The Toolkit Years
FAIR ships the picks and shovels of the deep-learning boom — detection, translation, search, simulation — and quietly racks up best-paper awards. If you did computer vision or NLP in these years, you ran Meta code.
faiss
Billion-scale vector similarity search, open-sourced. Years later it becomes the quiet backbone of the entire vector-database and RAG era.
fairseq
The sequence-modeling toolkit behind a decade of Meta breakthroughs — from convolutional translation to RoBERTa, BART, NLLB and wav2vec.
Mask R-CNN wins ICCV Best Paper
Kaiming He's elegant instance-segmentation architecture becomes the workhorse of an era — and the ancestor of everything 'Segment Anything'.
ELF OpenGo
An open reimplementation of AlphaZero-style Go that beats top professionals 20–0 — then gives the whole system away.
Habitat
A photorealistic 3D simulator where embodied agents learn to navigate thousands of times faster than real time.
Pluribus beats the poker pros
Six-player no-limit Texas hold'em falls — on about $150 of cloud compute. Science publishes it; the pros confirm it bluffs.
RoBERTa
'A Robustly Optimized BERT' — proof that training recipe matters as much as architecture. It tops GLUE and becomes the default encoder for years.
Codec Avatars
Reality Labs Research shows photoreal telepresence avatars — a decade-long bet that your VR self should look exactly like you.
Scaling Ambitions
The lab goes big: self-supervised learning, 175-billion-parameter models, 200-language translation, protein folding at planetary scale — and two very public lessons (OPT's brutally honest logbook, Galactica's three-day demo) about doing it in the open.
wav2vec 2.0
Speech recognition from raw audio with almost no labels — the foundation of everything multilingual that follows.
DINO: attention discovers objects
A self-supervised vision transformer whose attention maps segment objects nobody ever labeled. The rainbow lineage starts here.
Ego4D
Nearly 3,000 hours of first-person video from around the world at launch — grown to 3,670 by the 2022 public release — teaching AI to see life through human eyes.
OPT-175B and the logbook
A GPT-3-scale model released to researchers — along with its unvarnished training diary: 35 restarts, hardware failures almost daily. Radical transparency, warts and all.
No Language Left Behind
One model translating between 200 languages, including 55 African languages most systems ignored. Later published in Nature; its techniques serve 25 billion translations a day.
Make-A-Video
Text-to-video before it was normal — trained without a single paired text-video example.
Galactica lasts three days
A 120B model trained on scientific text meets the internet, hallucinates citations with confidence, and is pulled after 72 hours — a cautionary tale the whole field remembers.
CICERO plays Diplomacy
It negotiates with humans in free-form English across 40 online games, scores double the human average — and nobody calls it out as a bot during play. Science publishes it.
ESMFold maps 617M proteins
A protein language model folds the entire metagenomic dark matter of the planet in two weeks on 2,000 GPUs. Structure emerged from sequence alone.
The Open Revolution
On February 24, 2023, a research-only model called LLaMA changes everything — especially once it escapes. Meta leans in: open weights become strategy, SAM makes segmentation a commodity, and by late 2024 a downloadable model matches GPT-4.
LLaMA — and the leak
A 13B model that beats GPT-3's 175B rewrites the scaling playbook. Within a week the research-gated weights are on 4chan — and the open-weight LLM ecosystem is accidentally born.
Segment Anything
Click anything, cut out anything: one promptable model trained on 1.1 billion masks makes segmentation a solved commodity — and ships with a browser demo on day one.
DINOv2
Label-free visual features that beat supervised pipelines. NASA JPL later adopts them for Mars robots; the World Resources Institute maps forests with them.
Massively Multilingual Speech
Speech recognition and synthesis for 1,100+ languages — more than ten times the coverage of prior systems.
Llama 2: open for business
The most capable commercially usable weights yet, and the first Llama you could build a business on — launched arm-in-arm with Microsoft. 'Open weights you can build a business on' becomes Meta's official strategy.
AudioCraft & SeamlessM4T
MusicGen turns a sentence into a song; SeamlessM4T translates speech across ~100 languages — the Babel Fish gets real, and later lands in Nature.
Purple Llama
Llama Guard and CyberSecEval: safety tooling released as openly as the models it protects.
V-JEPA: learning by watching
Yann LeCun's world-model program bears fruit — video models that learn physics by prediction in representation space, not pixels.
Llama 3.1 405B
The first GPT-4-class model anyone can download: 405B parameters, 16,000 H100s, a 559-author paper — and Zuckerberg's essay 'Open Source AI Is the Path Forward.'
SAM 2 tracks video
One click on one frame follows an object through an entire video, in real time.
Movie Gen
Type a sentence, get a 16-second HD film with its own synchronized score — a 92-page report, 88 authors, ~100 million training videos.
The architecture bets
In one remarkable month FAIR publishes the Byte Latent Transformer (kill the tokenizer), Large Concept Models (predict ideas, not words), Coconut (reason in latent space) and Memory Layers — public bets on what comes after the transformer status quo.
The Whiplash Year
Llama 4 lands to mixed reviews and its 2-trillion-parameter teacher never ships. Zuckerberg responds with the most aggressive talent raid in tech history — $14.3B for half of Scale AI — and rebuilds everything as Meta Superintelligence Labs. Meanwhile FAIR quietly has one of its best research years ever.
Llama 4 — and a turning point
Scout arrives with a 10-million-token context window (Maverick gets 1M); the Behemoth teacher is previewed and repeatedly delayed. By LlamaCon, Llama has passed 1.2 billion downloads — but the frontier crown has slipped.
Meta Superintelligence Labs
$14.3B buys 49% of Scale AI — and its 28-year-old CEO Alexandr Wang becomes Meta's Chief AI Officer. MSL forms over FAIR, GenAI and product AI, and nine-figure offers start landing in OpenAI researchers' inboxes.
V-JEPA 2 controls robots zero-shot
Trained on a million hours of video and 62 hours of robot data, it picks up objects with arms it has never controlled, in labs it has never seen.
DINOv3
A 7B self-supervised vision model trained on 1.7 billion images, with smaller ViT and ConvNeXt variants distilled from it — one frozen backbone that works from Instagram photos to satellite passes over the Amazon.
Ray-Ban Display & the Neural Band
A decade of Reality Labs sEMG research — published in Nature that July — ships on wrists: glasses you control with imperceptible finger movements.
Code World Model
A 32B model that mentally simulates code execution — variable states, line by line — before writing a fix. Released with checkpoints from every training stage.
The reorg bites
Two months after the four-group split, 600 AI roles are cut in a single day — TBD Lab largely spared — even as superintelligence hiring continues. The message: smaller, denser, faster.
SAM 3, SAM 3D & Omnilingual ASR
Type 'yellow school bus' and SAM 3 masks every one, in video, in near real time; SAM 3D lifts single photos into 3D; Omnilingual ASR transcribes 1,600+ languages — hundreds for the first time ever.
Yann LeCun leaves
After twelve years, FAIR's founder announces his exit to chase world models at his own startup — which soon raises one of the largest seed rounds ever, $1.03B at a $3.5B valuation, before shipping anything. The JEPA program continues at FAIR without him.
SAM Audio
Segment Anything — for sound. Isolate any voice, instrument or noise from a mix by clicking, describing, or pointing at it on video.
The Muse Era
MSL's first models arrive under a new name — and a new strategy. Muse Spark is Meta's first closed frontier model; three months later Muse Glimmer throws the doors open again. Both things are true at once: Meta is done giving away the crown jewels, and determined to keep leading open source.
The infrastructure flex
MTIA 300 silicon deploys at scale with a 400/450/500 roadmap; the gigawatt Prometheus supercluster is slated to come online this year; Hyperion grows toward 5GW and $50B+.
Muse Spark
MSL's debut: a natively multimodal reasoning model with Instant, Thinking and Contemplating modes — the last orchestrating parallel reasoning agents. Closed weights, billions of users, and a comeback on the leaderboards.
Muse Image & Muse Video
An image model that reasons, self-refines and renders legible text — #2 on three Arena leaderboards at launch — plus a video model in preview. Every output carries an invisible, crop-proof Content Seal watermark.
Goodbye, Llama API
Meta folds its hosted Llama service into the new Meta Model API — the clearest sign the Llama era is closing.
Muse Code
Meta's first coding agent: terminal-based, multi-subagent, crash-proof via an append-only event log — powered by the new Muse Spark 1.2.
Muse Glimmer — the plot twist
A 30B open-weight agent model under Apache 2.0 — more permissive than any Llama license ever was — distilled from Spark 1.2 and small enough for a laptop. It ships alongside a public recommitment from Zuckerberg to open source, with an open Spark 1.2 promised next.
That is the story from inside Meta. For where it leaves Meta against everyone else — who the world actually downloads, whose licence lets you build, and the dimensions on which Meta now trails — see Meta against the frontier.
…and the sky keeps growing. Back to the Universe ✦
Or: untangle the family trees · see it by the numbers · test yourself