What World Models Are: Atlas, World Labs and the AMD Deal
World Labs defines its subject in one line: world models generate, reconstruct, and simulate any possible world. Its Atlas, released 2026-09-01, is the closest thing to a product so far — early access only, with no published price and no public API. On 2026-09-28 AMD announced an all-stock acquisition of World Labs.
Updated 2026-09-29
Four key points
The one-line definition
In World Labs' words: "World models generate, reconstruct, and simulate any possible world." Those three verbs are the bar; a generator that only does one of them is not one.
How Atlas is offered
The official page offers only "Request early access"; price and public API are unpublished, so no figures appear here.
The official deal figure
AMD's own release: an all-stock transaction valued at approximately $8.2 billion, expected to close by the end of 2026, subject to regulatory approvals and other customary closing conditions. The figure is self-reported by the acquirer, not a third-party valuation.
Reading images ≠ building worlds
What QCode sells today are multimodal language models — the official page lists image input for gpt-6-sol, for example. That reads images; it neither generates nor simulates a walkable 3D world.
How a world model differs from an LLM
In World Labs' own words the job is to generate, reconstruct and simulate worlds: the model is about how a world's state evolves under actions, whereas a language model is about the next token. The difference is in what is being modelled, not in size: one has to land in geometry, dynamics and pixels, the other in text. That is why Atlas is described as natively operating on text, images, video and 3D rather than as a bigger chat model. On QCode the two are currently different products: language models are callable, world models are not offered.
The year's dividing line: a world-model lab gets acquired
On 2026-09-28 AMD officially announced a definitive agreement to acquire World Labs: an all-stock transaction valued at approximately $8.2 billion, expected to close by the end of 2026, subject to regulatory approvals and other customary conditions. The same release states that Dr. Fei-Fei Li will join AMD as executive vice president and chief scientist. It is the first time the world-model route has been priced by a large acquisition rather than shown as a research demo.
Public timeline
On 2026-09-01 World Labs released Atlas, officially described as "our new omni world model for spatial intelligence", with the definition above and the vendor-stated ability to output up to 1 minute of video at 1440p; the only path offered is requesting early access.
On 2026-09-24 Sakana AI announced Jürgen Schmidhuber joining as Chief Scientific Advisor, quoting his line on the direction: "The future of intelligence is not just language; it is physical AI powered by World Models." That is a researcher's framing of a route, not a product fact.
On 2026-09-28 both companies published announcements the same day. AMD's release states an all-stock transaction valued at approximately $8.2 billion, expected to close by the end of 2026 subject to regulatory approvals, with Dr. Fei-Fei Li joining AMD as executive vice president and chief scientist; World Labs' own headline reads "World Labs has signed a definitive agreement to join AMD".
Confirmed vs unverified
Confirmed (official text)
Re-checked word for word on the official pages on 2026-09-29, three points are confirmed. One: the definition — "World models generate, reconstruct, and simulate any possible world". Two: Atlas is described as an omni model pretrained from scratch to natively operate on text, images, video and 3D, built as a multimodal autoregressive diffusion transformer. Three: the only access path shown is "Request early access", and World Labs says Atlas will power future versions of Marble and other products.
Unverified · not carried here
As of 2026-09-29 this page deliberately leaves out three things: release dates and specs for other labs' world models (Google DeepMind's Project Genie, NVIDIA's Cosmos and so on — we did not verify them on official pages, so they are not stated here); any price, public-API date or benchmark for Atlas, which remains unpublished; and claims that world models will replace language models. The unverified column looks empty not because there is no chatter, but because none of it met our sourcing bar.
Two concepts side by side
World model vs language model
By the official definition a world model must generate, reconstruct and simulate any possible world, modelling world state under actions; a language model outputs a text sequence. Atlas is officially described as an omni model natively operating on text, images, video and 3D.
Vendor statements vs verified facts
Vendor claim: Atlas outputs up to 1 minute of video at 1440p ("outputting up to 1 minute of video at 1440p"). Verified fact: Atlas shipped 2026-09-01, its only listed access path is requesting early access, and there is no published price or API. The two are labelled separately here; we neither recompute nor extrapolate the vendor's number.
What to look at if you want to try one
Only two routes can be verified today. Atlas is request-based: the official page offers "Request early access", with no published price list and no public API endpoint, so any claim about what Atlas costs or when it opens for calls currently has no primary source. The second: if the goal is image understanding or video input inside a product, today's multimodal language models already cover that, and a world model is not required.
Where QCode stands
Up front: QCode does not offer Atlas or any other world model, and this page will not invent a price or an endpoint for something we do not sell. What QCode does carry today is multimodal language models: the official model page lists gpt-6-sol as "Input modalities: text, image", and on the Claude side the only page verified word for word here is Sonnet 5.5, whose overview reads Text and images → text; the Opus 5.5 modality line was not verified word for word, so it is not claimed. Reading an image is a different thing from generating or simulating a 3D world you could walk into.
Questions
How does a world model differ from an LLM?
The modelled object differs. The official definition gives world models three verbs — generate, reconstruct, simulate any possible world — so the model has to carry how a world's state changes under actions; a language model targets the next token. Atlas is officially described as natively operating on text, images, video and 3D, not as a longer chat model.
Can I try Atlas for free?
The official page shows only "Request early access" — no published price and no public API endpoint — so neither "free" nor a price figure has a primary source yet, and this page does not guess. What is confirmed: released 2026-09-01, offered by request, and slated to power future versions of Marble and other World Labs products.
Why is AMD buying World Labs?
AMD's release describes World Labs as "an AI model and research lab led by AI pioneer Dr. Fei-Fei Li", states the all-stock terms of about $8.2 billion with an expected close by the end of 2026 subject to regulatory approval, and confirms Dr. Fei-Fei Li joining as executive vice president and chief scientist. On strategic intent this page restates only what the two announcements say and adds nothing.
Has the deal closed?
Not confirmed, and the two sides do not word it the same way: AMD's release says «entered into a definitive agreement» and gives «expected to close by the end of 2026, subject to regulatory approvals and other customary closing conditions», while World Labs writes «World Labs has signed a definitive agreement to join AMD». Together those support only «signed, not closed» — so this page says neither that the deal has completed nor that World Labs already sits inside AMD.
What about Google's Genie or NVIDIA's Cosmos?
This page does not cover them. We could not verify their dates or parameters on the vendors' own pages, and unverified material is not written here; second-hand figures and rumour are not our sources either. Should an official page and date be verified later, a background line would be written from it rather than copying someone else's numbers.
Will world models replace language models?
This page does not answer that — it is a prediction, not a verifiable fact. What can be checked is that both exist now: the official definition asks world models to generate, reconstruct and simulate, while what is callable and sold today are language models and multimodal language models (including on QCode). Schmidhuber's line quoted in Sakana's announcement (physical AI powered by World Models) is one researcher's judgement, so it sits in the timeline rather than in a conclusion.
Sources
World Labs' Atlas release page (definition, architecture and early access), AMD's official press release on acquiring World Labs, World Labs' own join-AMD announcement (dated September 28, 2026 at the top of the page), Sakana AI's announcement about Schmidhuber joining (2026-09-24), and this site's product status captured on 2026-09-29.
Start with what you can call
QCode does not sell world models; what is callable right now is multimodal language models such as gpt-6-sol, which lists image input (as of 2026-09-29).
Related
What recursive self-improvement means
The other line of the same 'next paradigm' question: models improving models, with the official Schmidhuber and Sakana wording.
GPT-6 Sol status tracker
One of the multimodal language models you can actually call: the official page lists image input.
Claude Sonnet 5.5: the full guide
The other end of this page's bridge: Anthropic's 5.5-family language models, their specs and prices — also not world models, but callable on QCode today.
QCode is not affiliated with World Labs or AMD. The definition, dates, access path and deal figures are taken verbatim from the two companies' own pages and press releases, as they stood when captured on 2026-09-29; capability numbers supplied by the vendor (for example 1 minute of video at 1440p) are labelled as self-reported and were not recomputed by us. QCode sells no world models and does not forecast any company's future product plans.