World Labs Atlas: Omni World Model for Spatial Intelligence

What shipped

World Labs — Fei-Fei Li's stealthy startup that raised $230M+ — dropped Atlas yesterday, and it's not another video generator. Atlas is an omni model pretrained from scratch on text, images, video, and 3D. The architecture is a multimodal autoregressive diffusion transformer: all inputs get merged into a shared spatial context grounded at 3D positions, and Atlas generates what comes next, staying 3D-consistent across everything it sees.

The capabilities span four domains:

Why it matters

World models have been theory for years. Atlas is the first production-grade example that unifies generation, reconstruction, and simulation in a single architecture — and it outperforms specialized models at their own games. Third-party raters preferred Atlas over MiniMax H3 (75%), Gemini Omni Flash (81%), and FLUX 3 (93%) for camera-controlled generation. On 3D reconstruction error, Atlas beats dedicated models without breaking a sweat.

The robotics angle is the sleeper hit here. Scanning a real space with a phone, then having Atlas simulate novel robot trajectories through it with photorealistic sensor data, collapses the sim-to-real gap dramatically. For anyone building embodied AI, this is the infrastructure piece that's been missing.

Early access is open via Typeform. If Atlas scales as promised, this is the year world models become a platform, not a paper.