World Labs unveils Atlas: a bet on spatial models, where the claims outrun the evidence
On 1 September 2026, World Labs — the startup co-founded in 2024 by Fei-Fei Li, formerly director of the Stanford AI Lab — used its official blog to introduce Atlas, described as “an omni model that we pretrained from scratch to natively operate on text, images, video, and 3D”. The architecture is a multimodal autoregressive diffusion transformer trained as a rectified flow model, in which inputs are anchored in a shared three-dimensional spatial context rather than stitched together by adapter layers between separate systems. The company, which raised one billion dollars in February 2026 in a round joined by Autodesk, NVIDIA and AMD, states that Atlas generates up to a minute of video at 1440p resolution, handling the camera trajectory through geometric inputs and returning output as point clouds and 3D Gaussian splats — the same representation used by Marble, the only product the company has had on the market since November 2025.
On camera-control comparisons, the official announcement cites tests run by third-party human raters showing a preference for Atlas of 75% over MiniMax H3, 81% over Gemini Omni Flash, 86% over Happy Horse 1.1, 93% over FLUX 3 and 94% over Seedance 2.5; the comparisons are chosen and run by World Labs, which hands the judgement to external human raters (“Third-party human raters judge which model better follows the intended camera path”) without disclosing how many raters took part, which prompts were used or what criteria applied. For three-dimensional reconstruction, the outlet The Decoder reports a stated median error of 25.3, the lowest figure among the models World Labs picked for the comparison. Discrepancies emerge over the input data, however: while the company blog speaks of an input volume ranging from a handful to more than a hundred images (with faithful reconstructions from two or three ground-level photos), trade-press analyses give a range from one to several dozen images.
There is no scientific paper, no model card, no arXiv listing, no model weights and no source code: the total parameter count, the composition of the training data, the training costs and the system's actual usefulness in fields such as robotics and simulation therefore remain unverifiable. Atlas is not a commercial launch but an early-access programme for partners selected through a sign-up form, with no indication of pricing, public APIs or timelines for general availability, even though World Labs states that “Atlas will power future versions of Marble and other products from World Labs”.
The shift from frame generators to world models shows where the industry is steering its resources: the challenge is no longer producing striking images, but measuring three-dimensional space. Until papers, model cards and weights are out, what we have of Atlas is what World Labs says about Atlas. The real measure will come from whoever uses it outside the demos — if and when access opens up.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- Read the official announcement at worldlabs.ai/blog/atlas in two passes, the second to pull the exact wording on the percentages, the early-access programme and the relationship with Marble: that is where the architecture, the minute at 1440p, the Gaussian splats and the five comparisons come from. Independently confirmed on The Decoder (which also reports the 25.3 median error and the company timeline) and on SiliconANGLE, both dated to the launch. The funding figure comes from TechCrunch's article of 18 February 2026. Verified that the announcement comes with no paper, no arXiv entry, no model card and no code: that is why the numbers remain self-reported. Dropped radiancefields.com as a primary source (it writes “MipMap H3” where the official post says “MiniMax H3”) and The Information because it is paywalled.
- Incertezze
- No figure can be verified from outside: there is no paper, no model card, no weights, no code and no independent benchmark, and the comparisons are chosen and run by World Labs itself, with a method — human raters' preference on adherence to the intended camera path — that is described but not documented as to the number of raters, the prompts or the criteria. Parameter count, training data and costs, pricing, APIs and a public availability date are all undisclosed. On the minimum and maximum number of input images, secondary sources diverge from the announcement. The date is 1 September 2026 according to the official post and SiliconANGLE, 2 September according to The Decoder. Real usefulness in robotics and simulation remains unmeasured outside the published demos.
- Perché pubblicarla
- It is the announcement of a frontier model in a category the site had not yet covered — world models — and it comes from a company funded by NVIDIA, AMD and Autodesk, with the whole hardware and CAD chain invested in the outcome. But the story matters above all for the gap between what is claimed and what can be checked: five percentage wins over opponents picked in-house, no paper, no weights, no price, invitation-only access. It is exactly the kind of announcement this site can tell well, keeping the facts (“World Labs states”) separate from the evidence (none, so far, from outside the lab).