What is a world model? The three kinds, and the one question that separates them

“World model” is now attached to at least three different kinds of system, and the differences between them decide whether a tool is any use to you. A model that reconstructs your apartment as a file you can open in Blender and a model that renders a playable dream you cannot save are both called world models, in the same sentence, by the same article.

This page sorts them out, and gives you a test you can apply to the next model that gets announced. Where a company reports its own results, it says so.

The one question that separates them

Skip the marketing copy and ask one thing:

When the model is finished, what do you have?

Three answers, three kinds of system:

1. A 3D scene you can keep — point clouds, Gaussian splats, a mesh. Files that open in other software and outlive the session.
2. A session you can play but not save — a world that stays consistent while you move through it in real time, and disappears when you close it.
3. A video file — pixels along one path, with no world behind them to revisit.

All three get called world models. This page argues that only the first two earn the name in any way that helps you choose a tool — the case for including the third is set out in its own section below. The rest of this page takes each in turn.

Kind 1: models that hand back 3D

These reconstruct or generate a space and give it to you as geometry. The space exists after the model stops running, which is what makes it usable as an asset, a game level, or a robot’s training environment.

Atlas, announced by World Labs on September 1, 2026, outputs point clouds and Gaussian splats alongside camera-controlled video (Dealroom). World Labs describes it as filling gaps to “turn point clouds into complete splat scenes that render on device.” Atlas is in early access for select partners, so for now this is a description of what it does, not something you can call.

Marble, the World Labs product that shipped first, is the one you can use. Its World API has been public since January 21, 2026 (World Labs) and exports Gaussian splats as .spz or .ply, plus a mesh and a fly-through video (Radiance Fields). As of September 7, 2026, among the seven models on this page it is the only one combining public developer access with 3D export — the full accounting is in world model APIs in 2026.

The practical signature of this group is that the output has a file extension. You can open it in a viewer, drop it into Unity or Unreal, hand it to a colleague. If you want to check what a Gaussian splat actually looks like, the free splat viewer on this site opens .ply and .spz files in the browser.

Kind 2: models you can play but cannot export

These maintain a consistent world in real time and let you move through it and change it. The world holds together for as long as the session lasts. Then it is gone.

Genie 3 from Google DeepMind runs at 720p and 24 fps, holds a scene consistent for several minutes, and accepts text prompts mid-session to change the weather or add objects (DeepMind). It reached the public through Project Genie for Google AI Ultra subscribers in the US in January 2026. There is no 3D asset export.

GWM Worlds 2 from Runway, announced September 3, 2026, sits in the same group: interactive, and a research preview reachable through a contact form rather than an API (Runway). Our GWM Worlds 2 guide covers what it does in detail.

This group is often described as more advanced than the first, and for interaction it is. But the two groups are answering different questions. Genie 3 answers “what would it be like to be in this place right now.” Marble answers “give me this place as a file.” Neither substitutes for the other, and the Atlas vs GWM Worlds 2 comparison is really a comparison between those two questions.

Kind 3: video models that get called world models

This is where most of the confusion is generated, because the label is applied by people writing about the models rather than by the people building them.

MiniMax H3 generates 4 to 15 second clips at native 2K with 24 fps and stereo audio (MarkTechPost). Sora 2 produced 15 to 25 second clips; its API closes on September 24, 2026 (OpenAI Help Center). Both are excellent at what they do, and neither leaves a world behind.

The argument for calling them world models goes: to generate physically plausible video, a model must have learned something about how objects occupy space and persist through time, so there is an implicit world model inside it. That argument is about the model’s internals. It says nothing about what you can do with the output.

The gap shows up the moment you need a second shot. Ask a video model for the same room from another angle and you get a different room — similar furniture, different geometry, nothing that lines up. There was never one world; there was a plausible sequence of frames. This is exactly the failure the first two groups are built to avoid, and it is why Atlas takes a camera path as geometry instead of as a sentence describing camera movement.

A useful way to phrase it: video models are judged on whether a clip looks right. World models are judged on whether two clips agree.

The three kinds, side by side

Model	Vendor	What you get back	Export 3D?	Access as of 2026-09-07
Atlas	World Labs	Video, point clouds, Gaussian splats	Yes	Early access, select partners
Marble	World Labs	Splats (.spz, .ply), GLB mesh	Yes	Public World API
Genie 3	Google DeepMind	A real-time session, 720p / 24 fps	No	Project Genie, US AI Ultra
GWM Worlds 2	Runway	An interactive session	No	Research preview
Odyssey-2	Odyssey	Interactive video	No	Public developer API
MiniMax H3	MiniMax	A video file, 4–15 s, 2K, with audio	No	Open weights and hosted APIs
Sora 2	OpenAI	A video file	No	API closes 2026-09-24

Rows one and two are kind 1, rows three to five are kind 2, rows six and seven are kind 3. Sources for every figure are listed at the end of this page.

Nearby terms that are not the same thing

Four labels sit close enough to cause trouble.

AI 3D model generators. Tripo, Meshy, Hunyuan3D and Meta’s SAM 3D turn a photo into one object as a mesh — a chair, a helmet, a statue — ready to place in a scene somebody else built. World models generate the scene: the room, the street, the landscape you move through. The search terms overlap badly (image to 3d mostly means the object kind), so it is worth being explicit about which you need. Object generators vs scene generators sorts out which tool fits which task, and why the output formats differ.

Gaussian splatting and NeRF. These are not categories of model; they are ways of representing a captured scene, and a world model may use them as its output format. Classical Gaussian-splatting pipelines also need dozens to hundreds of overlapping photos, where a generative world model is designed to work from a handful and invent the rest. The trade-offs are in Gaussian splatting vs NeRF.

Digital twins. A digital twin is bound to a specific real asset and usually to live data from it, and it is judged on measurement accuracy. A generated world is judged on plausibility. A world model can be a starting point for a twin, but a plausible reconstruction is not a twin.

Spatial intelligence. This is a capability claim, not a product category — World Labs uses it to describe what its models are reaching for (World Labs). It tells you the ambition, not what comes out the other end. Apply the same question: what do you have when it finishes?

Where the term came from

The phrase predates all of this. In 2018, David Ha and Jürgen Schmidhuber published a paper titled World Models (arXiv:1803.10122) describing an agent with three parts: a component that compresses each frame into a small vector, a recurrent component that predicts what the next compressed frame will be, and a small controller that acts on those two. The striking result was that the agent could be trained largely inside its own learned simulation of the environment rather than in the environment itself.

In that usage a world model is internal machinery, invisible to any user, whose purpose is to let an agent plan and practise. The output is a policy, not a picture.

Today’s usage keeps the core idea — a learned model of an environment — and points it outward. The output is meant for people: frames to watch, scenes to walk through, geometry to import. Both usages are legitimate and you will meet both in search results, which is worth knowing if you arrived here from a robotics or reinforcement-learning context and found a page about Gaussian splats.

Why the distinction is worth the trouble

Three places it changes a decision:

Buying. If your pipeline ends in Unity, Unreal or Blender, only kind 1 can complete it. A kind 2 model can be the most impressive demo you have seen this year and still be unable to give you a single file. Deciding on the demo instead of the output format is the most common expensive mistake in this category.

Budgeting. The three kinds are priced on different units. Kind 1 tends to charge per world: a Marble world is a flat 1,500 credits, roughly $1.20, no matter how long you spend inside it (World Labs docs). Kind 3 charges per second of output. Comparing a per-world price to a per-second price without noticing the unit produces numbers that are simply wrong. Our pricing breakdown works through this.

Reading benchmarks. A comparison between a video model and a reconstruction model is usually measuring one of them on a task it was not built for. When you see one model reported as beating another, check first whether they belong to the same kind — and then check who ran the evaluation. World Labs’ own Atlas numbers are self-reported human preference scores with no published paper, which we unpack in Atlas benchmarks explained.

Checking any new model in about a minute

When the next one is announced, four questions settle which kind it is:

1. What formats does it export? Named formats (.ply, .spz, .glb, mesh, point cloud) mean kind 1. “Video” or “MP4” alone means kind 3. Silence usually means kind 2.
2. Can you re-enter the same world later, or only regenerate a similar one? Re-entering means there is a persistent world. Regenerating means there is not.
3. How is the camera specified? As geometry — a path, a pose, a trajectory — or as words in a prompt? Geometry indicates the model has an actual 3D notion of the scene.
4. Is there a public API, and what does its response body contain? This is the question that cannot be answered with marketing copy. As of September 7, 2026, four of the seven models on this page have no public API at all.

If you cannot answer question 1 from the announcement, that is itself the answer for planning purposes: assume no export until a source says otherwise.

Where to go next

• What is the Atlas world model? — the kind 1 model this site covers in depth
• World model APIs in 2026 — which ones you can actually call from code
• Atlas vs Marble — the shipping product versus the new model, both kind 1
• Compare world models — head-to-head pages across all three kinds
• Who is World Labs? — the company behind Atlas and Marble