World Labs released Atlas, a world model it says was pretrained from scratch to operate natively on text, images, video, and 3D data. The company describes Atlas as a multimodal autoregressive diffusion transformer that folds every input into one shared spatial context, then generates what comes next while staying geometrically consistent with everything it has already seen. Marble, the spatial-generation product World Labs already sells, is where the company says the model lands next.
Atlas is pitched as one model doing three jobs that are usually handled by separate systems: world generation, 3D reconstruction, and simulation. According to World Labs, Atlas can generate up to a minute of camera-controlled video at 1440p from a handful of reference images, rebuild a real scene in explicit 3D from two or three photographs, and simulate how a robot’s sensors would see a space as it moves through it.
That last capability is the one worth watching. A single model that reconstructs a real environment and then simulates an agent moving through it, in the same shared representation, is aimed squarely at robotics and agent training, where building a usable simulated environment has always been the expensive, manual part of the pipeline. World Labs says Atlas can turn a cell phone video of a room into a simulator that varies object positions, lighting, and robot motion for training data, without a scanning rig.
World Labs backs the release with head-to-head evaluations rather than a single leaderboard score. For camera-controlled generation, the company had third-party human raters compare Atlas against five video models: MiniMax H3, Gemini Omni Flash, Happy Horse 1.1, FLUX 3, and Seedance 2.5. Raters preferred Atlas over each of them, from 75 percent against MiniMax H3 up to 94 percent against Seedance 2.5, with the margin widening as camera paths got more complex. World Labs also says Atlas beats specialized open-source 3D reconstruction models on standard benchmarks, though it has not published a comparable independent chart for that claim.
Those numbers come from World Labs’ own testing, not an outside benchmark body. The rival models it tested against had their camera instructions typed as text prompts, because Atlas is the only one of them that takes camera geometry directly rather than as description. World Labs acknowledges that better prompt engineering might narrow the gap for those systems, a caveat that applies to a comparison a vendor designs and runs against its own product.
The company frames Atlas’s improvement as a scaling story. It trained a series of increasingly large versions during development and says each jump in compute unlocked new capabilities, a trend it expects to continue. World Labs has not disclosed model size, training data volume, or compute figures to substantiate that claim beyond the assertion itself.
Atlas is not broadly available. World Labs is opening early access to select partners who request it, with no stated timeline for a general release. The company did not disclose pricing, inference cost, or hardware requirements for running the model at scale.
For teams building robotics or agent-training pipelines, Atlas is worth a pilot request now, before broader availability turns evaluation slots into a queue. A model that generates and simulates from the same representation could cut the cost of building training environments faster than incremental gains on any single reconstruction benchmark would.
World Labs detailed the Atlas world model in a post on its own company blog.
