
World Labs builds Large World Models (LWMs) for spatial intelligence and 3D world generation.
World Labs' commercial lineup is anchored by Marble, a generative 3D world model that turns text, images, videos, or 360-degree panoramas into spatially consistent, persistent, and editable 3D environments. Marble is offered through a web studio and a paid World API on platform.worldlabs.ai that generates navigable 3D worlds from a single API call.
In September 2026 the company introduced Atlas, its next-generation omni world model for spatial intelligence. Pretrained from scratch as a multimodal autoregressive diffusion transformer, Atlas generates image and video frames along designed camera paths, reconstructs real scenes from sparse views into depth maps, point clouds, and 3D Gaussian splats, and targets architecture, filmmaking, and robotics-simulation workflows.
Spatial intelligence is emerging as the next major AI frontier after language models, with applications spanning gaming, robotics, simulation, digital twins, and immersive content.
World Labs competes in a growing field that includes NVIDIA's Cosmos platform, Google DeepMind's Genie family of world models, and startups such as Odyssey and Decart, all racing to build generative systems that understand and simulate the physical world.
World Labs pairs a founding team of computer-vision and graphics researchers—including a co-creator of NeRF—with an output-first product strategy: Marble worlds are persistent, editable, and exportable in standard 2D and 3D formats rather than streamed as video.
The Atlas model consolidates generation, reconstruction, and simulation in a single pretrained multimodal model with camera-path control, while the SceniX acquisition adds real-to-sim-to-real tooling for robotics customers, extending the company's position beyond creative workflows into physical-AI applications.