The Lab
The lab traces back to FAIR (Facebook AI Research), founded in 2013 under Yann LeCun. After a decade of mostly academic output, Meta pivoted to open-weights AI in 2023 with the original Llama release. Llama 4 in 2026 is the open-weights frontier — weights downloadable, license commercially permissive, available on every cloud, every laptop, every phone with enough RAM.
The unusual strategy: Mark Zuckerberg's bet is that open models commoditise the layer above the chip. If anyone can run a frontier-class model, the value moves to apps, distribution, and hardware — three layers Meta already dominates. The Reality Labs hardware roadmap (smart glasses, AR/VR) needs a model running locally; making that model open accelerates the ecosystem Meta wants to capture.
The catch is twofold. First, open weights are easy to weaponise — every release sparks a fresh debate about misuse and a fresh round of regulatory attention. Second, "open" doesn't mean open development. Meta picks the training data and the alignment, releases finished weights, but the lab itself is closed. You can run Llama. You can't really see how it was built.
Recent Coverage
Models
The flagship open-weights model. 405B mixture-of-experts, weights downloadable, license commercially permissive. The default starting point for serious self-hosters.
Best for
The smaller sibling. Runs on a single H100 at full precision, on consumer hardware quantised. The model researchers actually run locally.
Best for
Native multimodal with audio in/out and video in. Built for the Reality Labs hardware roadmap. Real-time conversation latency under 300ms.
Best for
When to Pick Meta AI · When to Pick Someone Else
Each lab in the atlas comes with its own positioning, model line, and use cases. The point of organising the AI landscape by lab is that the answer to "which model should I use" almost always starts with "which lab is closest to what I'm trying to do."
