World Labs Unveils Atlas, a Model That Thinks in 3D

World Labs, the startup founded by the Stanford researcher Fei-Fei Li, has unveiled Atlas, a model it describes as an "omni world model" for spatial intelligence. Announced on September 1, Atlas is built to work not only with words but with space itself, generating images and video that stay consistent in three dimensions as a virtual camera moves through a scene.

Most of the best-known AI models are language models at heart, with images and video bolted on afterwards. Atlas takes a different route. World Labs says it was pretrained from scratch to operate natively across text, images, video and 3D at once, folding every input into a shared spatial context and then generating what should come next. The technical label is a multimodal autoregressive diffusion transformer. In plainer terms, it treats a scene as a coherent three-dimensional space rather than a flat picture.

What it can do

The company says Atlas can produce camera-controlled imagery and video at resolutions up to 1440p and for stretches of up to a minute, holding the geometry of a scene steady as the viewpoint shifts. That steadiness is the hard part. Video generators have become remarkably good at short, striking clips, yet they tend to lose track of where objects sit once the camera turns. Keeping a world stable in 3D is precisely what Li has argued for years is the next real frontier, the move she calls going "from words to worlds."

Ambition ahead of detail

For now the reveal is more showcase than open door. Atlas is entering early access with a small group of selected partners, and World Labs has held back the research paper, the pricing and the names of those partners. That leaves outside researchers with polished demonstrations and few hard numbers to test them against, which is a fair reason to keep some skepticism in reserve.

The company has room to be patient. World Labs has raised roughly 1.2 billion dollars since 2024, including a 1 billion dollar round earlier this year. That kind of backing buys the time to turn an eye-catching demonstration into something people can build with, and it signals how seriously investors are taking the idea that the next leap in AI may be less about language and more about space.