World Labs, whose co-founders include Fei-Fei Li, has unveiled Atlas, a world model developed for spatial intelligence. Described by the company as an “omni model,” Atlas is a multimodal autoregressive diffusion transformer pretrained from scratch that processes text, images, video, and 3D data within the same architecture. By combining inputs in a shared spatial context, the model can predict unseen parts of scenes.
Atlas accepts camera positions and movement paths directly as inputs and can generate videos at 1440p resolution, lasting up to one minute, using one to six reference images and specified paths. It can generate novel viewpoints, 3D geometry, point clouds, and “3D Gaussian splat” outputs from single or multiple images. Although World Labs says it outperforms some open-source models in 3D reconstruction from sparse images, the results have not been independently verified.
The model can create robotics simulation environments from phone footage. Atlas will serve as the foundation for Marble and other products and is currently available in early access to selected partners. Pricing, the date of general availability, model size, and training compute capacity have not been disclosed. World Labs was founded in 2024 and aims to develop systems that understand the physical world through its three-dimensional relationships.
Why it matters
Atlas’s key distinction is that it aims not only to process different types of data together, but also to evaluate them within the spatial relationships of a scene; this approach could make it possible to move beyond two-dimensional images in the production of digital content about the physical world. Its ability to generate new viewpoints, 3D geometry, and robotics simulation environments ties the technology to the modeling of physical environments, rather than making it merely a video-generation tool. However, the scope of its performance claims remains unclear, as some of World Labs’ comparative results have not yet been independently verified. Atlas’s limited availability to selected partners does not currently show how applicable this capability is across different use cases. The lack of disclosed information on pricing, the date of general availability, and technical scale also creates uncertainty about accessibility and resource requirements.
Background
World is not a new name in the FikirPilot archive: we published a news article mentioning the name in the past 90 days; that article was dated August 25, 2026.