World Labs Atlas
Vintage: 2026-09. Primary evidence is World Labs' official post Atlas: A World Model for Spatial Intelligence dated September 1, 2026, as hydrated in 2026-09-02-grok-com-ai-news-digest-2026-09-02-fable-5-1-astra-atlas-g20. Architecture and output claims are a snapshot of that issuer page.
World Labs Atlas
One-line summary: World Labs' September 1, 2026 "omni" world model — text, images, video, and 3D; multimodal autoregressive diffusion transformer; camera-controlled generation up to 1 minute at 1440p. Recap/AI Weekly 81–93% head-to-head scores were not on the issuer page retrieved.
What it is
An "omni" model pretrained from scratch to natively operate on text, images, video, and 3D. Architecture: multimodal autoregressive diffusion transformer; inputs share a spatial context. Official page: camera-controlled generation "outputting up to 1 minute of video at 1440p"; also spatial reconstruction, space-time simulation (including Real-to-Sim for robotics), and image / 360 generation. Atlas is entering early access with select partners and "will power future versions of Marble."
The Grok.com digest and AI Weekly chip that pointed at Atlas are pointers, not grain.
Why it matters to this thread
Multimodality and world models are in-scope here (avatar/video generation-as-product lives in threads/ai-video-generation; this page is the spatial-intelligence / world-model release). First dated official Atlas post the thread has as a citable source.
Key facts (from 2026-09-02-grok-com-ai-news-digest-2026-09-02-fable-5-1-astra-atlas-g20)
- From the same source (World Labs post, September 1, 2026): omni model on text, images, video, and 3D; multimodal autoregressive diffusion transformer with a shared spatial context.
- From the same source: camera-controlled generation up to 1 minute of video at 1440p.
- From the same source: spatial reconstruction; space-time simulation including Real-to-Sim for robotics; image / 360 generation.
- From the same source: early access with select partners; will power future versions of Marble.
- From the same source: official page claims Atlas outperforms specialized models on camera-conditioned generation (third-party human raters) and sparse-view 3D reconstruction — without the recap's 81–93% table.
What this source does not establish
- Recap / AI Weekly "81–93% head-to-head" is recap/aggregator-only. Those percentages were not present on the World Labs page retrieved. Do not file them as issuer fact.
- No official World Labs X permalink.
WorldLabsAIlookup failed (suspended).@worldlabsresolved to an unrelated London innovation platform — not cited. - Not a fetched eval table and not a generally available product.
- No ticker, 8-K, or stock-market tag.
Open questions
- Where (if anywhere) does World Labs publish the 81–93% head-to-head table the recap repeated?
- What "Marble" versions Atlas will power, and on what schedule.