A film of a kitchen shows you the kitchen. A world model is supposed to let you walk in and open the fridge. Three organisations now sell or show something under that name, and they do not agree on what the name covers. World Labs wrote in June 2026 that the phrase is “one of the most important and most overloaded terms in AI today”. We read how Google DeepMind, World Labs and Nvidia each define world models on their own pages, what each has released, and what Hugging Face lists for anyone who wants the files.
What is a world model, according to the companies building one?
Google DeepMind gives its definition in the Genie 3 announcement of 5 August 2025. World models, it says, are AI systems that use their understanding of the world to simulate aspects of it, so that an agent can predict how an environment will change and how its own actions will affect it. The same post calls world models a stepping stone on the path to artificial general intelligence. That is DeepMind’s claim.
Nvidia’s glossary entry on world models is shorter. It calls them “AI tools that understand the dynamics of the real world”, physics and spatial properties included, and says they take text, image, video, sound and movement as input to predict what happens next. Nvidia’s product term is world foundation model: a large pretrained model that a developer then trains further on their own data.
World Labs is the most candid about the muddle. Its June post traces the technical meaning to reinforcement learning, where an agent acts, the state of the world changes, and the agent only ever sees observations of that state. From that loop it sorts today’s products into three kinds. A renderer outputs pixels for human eyes. A simulator outputs state, meaning geometry and physics that a program can compute on. A planner outputs actions for a robot.
World Labs then files Genie 3 under renderers, next to its own research preview RTFM, and says a renderer has no explicit understanding of three-dimensional structure. That is one company’s reading of a competitor’s product. DeepMind’s post lists modelling the physical properties of the world among Genie 3’s capabilities.
How are world models different from a language model or a video generator?
The difference from a language model is the material. In World Labs’ words, a language model learns the statistical structure of text and a world model learns the statistical structure of space and time. The machinery overlaps. Nvidia’s glossary says an autoregressive world model predicts the next visual state the way a language model predicts the next word. Nvidia’s model card for Cosmos 3 describes two transformer towers, and one of them generates text by standard next-token decoding.
The difference from a video generator is interaction. DeepMind says Genie 3 generates each frame in response to the user’s inputs as they arrive, and has to refer back to what it generated earlier. The post calls that a harder problem than generating an entire video, because inaccuracies accumulate. Its figures for Genie 3 are 24 frames per second at 720p, a world that stays consistent for a few minutes, and visual memory that reaches back one minute.
That line is thin, and the makers’ own pages show it. Nvidia’s Cosmos page lists synthetic video data among the uses of Cosmos 3 and describes video generation as imagination for training physical AI. World Labs’ taxonomy counts a text-to-video model as a renderer, which is one of its three kinds of world model. By that taxonomy the model behind our LTX-Video test in ComfyUI, which gave us a blurry 3.9-second clip, would qualify.
The simulator kind is the one with a long history on ordinary computers. Nvidia’s glossary says the traditional approach had engineers program the physical rules by hand, which was precise in narrow conditions and useless outside them. MuJoCo, the physics engine DeepMind maintains for robot simulation, belongs to that older family, and we measured MuJoCo on an M4 Pro at about 63,000 physics steps a second on one core. A learned world model is a bet that the rules can come from data instead.
What have DeepMind, World Labs and Nvidia released as world models?
DeepMind announced Genie 3 as a limited research preview for a small cohort of academics and creators. On 29 January 2026 Google opened Project Genie, a prototype web app powered by Genie 3, to Google AI Ultra subscribers in the United States, with generations limited to 60 seconds. The Genie 3 post does not mention weights or a checkpoint, and a search of Google’s account on Hugging Face for Genie returned nothing when we ran it on 8 October 2026.
World Labs made Marble generally available on 12 November 2025. Marble builds 3D worlds from text, images, video or coarse 3D layouts and exports them as Gaussian splats, meshes or videos. The company opened a public World API on 21 January 2026 and introduced a second model, Atlas, on 1 September 2026, with early access by request for select partners. Its organisation page on Hugging Face lists no public models. On 28 September 2026 World Labs said it had signed an agreement to join AMD, expected to close by the end of 2026, and the announcement commits the combined company to widely accessible open models. Until one appears, Marble and Atlas run on World Labs’ servers.
Nvidia is the only one of the three with weights to download. Its Cosmos page says the Cosmos world foundation models are available under the OpenMDW1.1 license from the Linux Foundation. The model card for Cosmos3-Edge on Hugging Face gives a release date of 20 July 2026 and says the model is ready for commercial and non-commercial use. It lists robotics, autonomous vehicles and smart spaces as the use case.
Nvidia’s pitch to those customers is data. The robots that could record it are still few: UBTech reported 921 full-size humanoids sold in the first half of 2026, in the stock exchange filings we read. Recorded data is hard to get hold of too. When we looked for the dataset promised at the World Humanoid Robot Games, more than 2,500 hours by the organisers’ account, we could not find an hour of it to download a month later. Nvidia’s answer is generated video, which its Cosmos page says is not constrained by what has been physically captured.
Can you run a world model on your own machine?
For Genie 3, Marble and Atlas the answer today is no, because there is nothing to download. For Cosmos 3 there are files, and Hugging Face lists their sizes.
| Model | Parameters, per Nvidia | Repository size listed | Transformer folder listed |
|---|---|---|---|
| Cosmos3-Edge | 4B | 9.18 GB | 6.74 GB |
| Cosmos3-Nano | 16B | 35 GB | 30.3 GB |
| Cosmos3-Super | 64B | 133 GB | 128 GB |
*Source: Nvidia’s model cards and the Hugging Face file listings for each repository, read 8 October 2026.*
A file size is not a memory requirement, and we will not turn one into the other. We have not run any Cosmos model, so we have no speed or memory figure to report. What the cards do say is narrower than the word open suggests. Supported hardware is listed as NVIDIA Ampere, Hopper and Blackwell, and the operating system as Linux. The Cosmos3-Edge card adds: “We have not tested on other operating systems.” Only BF16 precision is tested, and the card says other precisions are not officially supported. The card does not mention macOS. Our test machine is a Mac, which puts it outside the set Nvidia says it tested.
Cosmos3-Edge, the smallest, generates video at 256p and 480p, between 50 and 150 frames per request, by its own card.
World Labs’ June taxonomy separates a renderer from a simulator, and its September announcement promises open models. If Atlas weights are published after the AMD deal closes, there will be a second downloadable model to set beside Cosmos 3, from a company that has written down what it thinks the term should mean. Its files will show which of the three kinds it shipped.
