About this field guide

01What this site is

metaai·lightalo is a field guide to everything Meta's AI labs have shipped — from the founding of FAIR in 2013 through PyTorch, Llama, Segment Anything, DINO, Seamless and the Muse era. It is written and maintained by one person, as a lightalo side project. It is not affiliated with, endorsed by, or sponsored by Meta, and nobody at Meta reviewed it.

One page looks outward. Frontier sets Meta beside OpenAI, Google DeepMind, Anthropic and the open-weight challengers, because a field guide that only ever measured Meta against itself would flatter it by construction. That page is deliberately the least comfortable one here: it lists where Meta trails, and it prints the ways its own comparisons could still mislead you.

It exists because this body of work is remarkable and surprisingly hard to see whole. The papers are scattered across arXiv, the code across hundreds of repositories, the announcements across years of blog posts — and much of it was released openly, for anyone to use. We wanted one place where a curious newcomer and a working researcher could both stand back and take in the shape of it. The tone is admiring on purpose; the facts are checked on purpose too.

02How it was built

The catalog began as a research sweep over primary sources: arXiv papers, GitHub repositories, ai.meta.com and about.fb.com announcements, plus reputable press for the facts that only exist in news coverage. Roughly 140 sourced entries were distilled into 119 projects, each placed in one of 12 constellations — The Muse Era, Language & LLMs, Vision, World Models & Embodied AI, Speech & Sound, Generative Media, AI for Science, Open Source Infra, Games & Strategy, Safety & Trust, On-Device & Silicon, and Reality Labs Research — and joined by 78 lineage links showing what grew out of what.

Every entry then went through an adversarial fact-check pass: a second reader was asked to find anything wrong rather than to confirm what was right. That pass corrected 33 claims — dates, model sizes, licence terms, who led what — before launch.

The Lab runs real models, entirely offline. Large models split into a heavy understanding half and a light interactive half, so we ran the heavy half once on a Raspberry Pi 5 and shipped the light half to your browser:

• Segment Anything — the real, 8-bit quantised SAM mask decoder runs in WebAssembly (via a self-hosted onnxruntime-web), against image embeddings precomputed on the Pi.
• DINOv2 — PCA feature rainbows and patch correspondences precomputed from dinov2_vits14; the hover matching is plain JavaScript cosine similarity. On the your-photo page the whole network runs in your browser (23 MB, 8-bit) and the PCA is computed in JavaScript — nothing is uploaded.
• OPT's tokenizer — the byte-level BPE behind OPT-175B, reimplemented in JavaScript and checked token-for-token against the reference implementation (Tokens).
• NLLB-200 — our sentences translated into all 202 FLORES-200 languages with CTranslate2 on the Pi.
• MusicGen — ten clips composed by musicgen-small, also on the Pi.

The whole site is static HTML, CSS and JavaScript: no build step, no frameworks, no analytics, no trackers, no cookies. Nothing you click here is sent anywhere.

03Sources & verification

Every star links to its primary sources. Open any project in the galaxy and the panel lists the paper, repository and announcement it was built from — they are the authority, not us. The catalog is a single file, /data/projects.json, which you are welcome to read, cite and build on.

What you may do with it. The facts here are not ours — dates, parameter counts and licence terms belong to whoever published them, and we claim nothing over them. What we did make is the compilation: the selection of 119 projects, the constellations, the lineage edges, the taglines and the written summaries. That compilation, and the JSON it lives in, is offered under CC BY 4.0 — use it, remix it, feed it to something, just say where it came from. The demo photographs are our own AI-generated images under the same terms. Model weights and outputs stay under their own licences, listed below.

Take it away. The catalog is published in three shapes, all under the same licence: catalog.csv is one flat row per project and opens in a spreadsheet; catalog-lineage.csv is the 78 lineage edges as parent/child pairs; projects.json is everything, exactly as this site renders it. The open_release column means weights or code were published, which is a weaker claim than an open-source licence — see the corrections for why that distinction is spelled out everywhere here.

How to cite it. There is a permalink for every project at /p/<id>/. A reasonable citation:

metaai·lightalo, "The Meta AI Universe" (an unofficial field guide),
  https://metaai.lightalo.com/, catalog verified 3 September 2026.
  Retrieved <date>.

If you are citing a specific claim, cite the primary source we link rather than us — that is what the links are for, and it is what we would do.

The “verified 3 Sep 2026” stamp in our footers means that on that date every fact in the catalog was re-read against its cited source, and that every link on the Live Demos page was loaded and confirmed alive. It is a snapshot, not a promise: Meta ships quickly and links rot. When something changes, the stamp changes with it.

Some numbers on this site are not ours and are not frozen. The stars, citations and downloads shown on each project and on the Numbers page are read straight from the GitHub, Semantic Scholar and Hugging Face APIs by a job that runs every Monday, and each carries the date it was read. They are attention and usage proxies, not achievements: citation indexes lag, download counts include automated traffic, and a starred repository is not a used one. Several catalog entries share one repository, so the figures are stored per repository and never added up per project.

How to read the badges on each star:

04Licenses & attribution

The Lab's models belong to their creators; we use them within their licences and say so on every page:

This site carries no advertising and sells nothing. The demo photographs are our own AI-generated images, made for these demos. Project names and marks belong to their owners; they appear only to identify the work being described, and no Meta logos or trade dress are used. The in-browser runtime is onnxruntime-web (MIT), served from this domain.

05Corrections

Fixes land in the catalog and refresh the verified date.

06Start here

Two ways in, depending on how much of this story you already know.