World Labs ships an omni world model

World Labs, the spatial-intelligence startup founded by Fei-Fei Li, unveiled Atlas on September 1: a multimodal autoregressive diffusion transformer pretrained from scratch to work natively across text, images, video, camera poses, and 3D depth maps inside one shared spatial context. World Labs calls it an “omni world model,” meaning the same system can generate a 3D scene, reconstruct one from partial input, or simulate how a space behaves, rather than needing separate specialized models for each task.

Atlas can generate 3D-consistent video at up to 1440p resolution and up to one minute long, and outputs explicit 3D geometry, not just flat video frames, so a generated scene can be walked through or queried for depth rather than only viewed. Training from scratch, rather than fine-tuning on top of an existing large language model, is what World Labs says lets it ground text, image, and video in one consistent 3D space instead of stitching separate modalities together after the fact.

The launch is early access, not a general API release, so there is nothing to integrate into a product yet. The signal is strategic: World Labs was founded specifically to pursue spatial intelligence as a distinct research bet from the LLM scaling race most labs are running, and Atlas is its first concrete model built on that bet. It arrives alongside a broader push into world models for physical tasks, including Nvidia’s own world-model work for robotics and a wave of large robotics funding rounds betting spatial reasoning is the next frontier past text and image generation. If you work in robotics, game or simulation tooling, or any pipeline that needs 3D scene generation rather than 2D image generation, this is the model to watch for a public API, not one to plan around yet.