GWM Worlds 2: Runway's playable world model explained

GWM Worlds 2 is Runway’s General World Model, announced on September 3, 2026. It generates a world you can enter and steer: one continuous 720p video stream at 24 frames per second, with audio at 48,000 Hz, responding to your inputs as you explore (Runway). Runway’s claim for it is not resolution but generality of control, that it is the first world model to generalise to arbitrary actions rather than a fixed action set or navigation alone.

This page is an independent summary of what has actually been published, with sources and dates. It is not affiliated with Runway.

What GWM Worlds 2 is

Architecturally it is an autoregressive diffusion model that generates video and audio together, built on top of Runway’s foundational audio-visual generation model. Autoregressive means each step is produced conditioned on what came before, which is what allows the world to stay coherent as you move through it rather than resetting between clips.

The practical consequence is a session rather than a render. You define a world, you enter it, and it keeps going. There is no preset length: the world continues from each new input instead of following a fixed clip or script.

It is the successor to GWM-1, which Runway released on December 11, 2025 as its first world model (TechCrunch). GWM-1 established the general-world-model line; Worlds 2 is where it becomes playable in real time.

How WorldPrompt works

The control format Runway calls WorldPrompt is the most interesting published detail, and the part least covered elsewhere. It splits a world into two kinds of state: what persists, and what changes.

The persistent half is a genesis prompt. It describes the environment, the subjects in it with their attributes and their voices, and the behavioural laws the world obeys, gravity, collisions, camera perspective. It is anchored by a first frame, which gives the model visual grounding for everything the text describes. Once set, this is the world’s constitution: it does not change as the session runs.

The changing half is what you do inside it. You steer with text actions addressed to any subject or to the scene itself, alongside continuous camera motion. Telling a character to climb is an action; telling the scene to darken is an action; moving the camera is a separate, continuous channel.

Underneath, each autoregressive step is conditioned on three kinds of context: the global context, meaning the genesis prompt and the first frame; the current frame’s inputs, meaning the camera input and the text actions spanning that frame; and the past generated frames, cached in a sliding window.

That third element is where the practical limits will live. A sliding window is finite, so how much history the model can hold determines how long a world stays consistent when you leave a room and come back. Nothing published states the window size, so treat long-session consistency as untested rather than claimed.

Specifications

	GWM Worlds 2
Video	720p, one continuous stream
Frame rate	24 fps, real time
Audio	48,000 Hz, generated with the video
Session length	No preset length
Control	Genesis prompt and first frame, then text actions plus continuous camera motion
Output files	None; frames only

Sources: Runway, Runway on X.

The 720p and 24 fps figures are the same as Genie 3’s, and that is not a coincidence. Both models have to deliver a frame before a human notices latency, and that budget is what caps resolution. Offline models such as Atlas run at up to 1440p precisely because they are not racing a human’s perception.

What it does not do

Three absences matter more than any of the specifications above.

No 3D export. GWM Worlds 2 produces video and audio frames. There is no mesh, point cloud or Gaussian splat output in the published material. Nothing survives the session as a file you can import into Unity, Unreal or Blender. If that is your requirement, the honest recommendation is to look at splat-producing models instead, starting with Atlas and the exportable worlds covered in the game development guide.

No public API. Nothing in the announcement describes developer access, endpoints or SDKs.

No published pricing. There is no per-second rate, no credit cost, no subscription tier that includes it. The access position is covered in detail in GWM Worlds 2 pricing and access.

Why arbitrary actions is the real claim

The interactive world models covered on this site before it, Genie 3 and Odyssey-2, accepted a limited vocabulary. Genie 3 lets you navigate and inject promptable world events. Earlier game-trained models accepted controller inputs mapped to a fixed action set. In both cases the space of things you can ask for was decided in advance.

Runway’s claim is that GWM Worlds 2 removes that ceiling: subjects and the scene can be directed to take arbitrary actions, expressed in text. Movement like running, climbing and leaping is the demonstrated case, but the format extends to gestures, object interaction, speech and sound.

If it holds outside curated demos, the category changes. A fixed action set gives you a playable environment. An open action set gives you something closer to a simulator you can interrogate, which is why Runway frames the model as a foundation for interactive entertainment, virtual characters, robotics and embodied-agent simulation rather than for video production.

The caveat is that generalisation claims are exactly the kind that degrade under adversarial use, and no independent party has tested this one. Treat it as promising and unmeasured.

How it compares to the other world models

Against Atlas (World Labs). The cleanest contrast in the field right now, and the two were announced two days apart. Atlas generates offline at up to 1440p and exports point clouds and Gaussian splats; GWM Worlds 2 streams at 720p in real time with audio and exports nothing. One gives you a file, the other gives you a place. Full breakdown in Atlas vs GWM Worlds 2.

Against Genie 3 (Google DeepMind). The closest technical sibling: same resolution, same frame rate, same real-time premise. GWM Worlds 2 adds audio and an open action vocabulary. Genie 3 has the decisive practical advantage of actually being reachable, through Project Genie for Google AI Ultra subscribers in the US since January 29, 2026. Detailed in GWM Worlds 2 vs Genie 3.

Against Marble (World Labs). Not really a comparison, more a division of labour. Marble is the only world model in this group with a live public API and published pricing, at roughly $1.20 per world, and it exports splats. If you need to ship something this quarter, see the Marble API guide and the pricing page.

Who it is for

• Interactive entertainment. The strongest fit by some distance. Audio plus open actions is what a playable world needs, and none of the models compared on this site currently offers both.
• Virtual characters. Subjects carry attributes and voices in the genesis prompt and can be addressed individually, which is unusual.
• Embodied-agent research. Generated environments with an open action space give volume and diversity that captured environments cannot.
• Not for asset pipelines. No export means no handoff to a DCC tool or a game engine.
• Not for reconstruction. GWM Worlds 2 generates worlds; it does not rebuild real places from photographs. That is what Atlas and Marble do, as covered in the robotics real-to-sim guide.

Access status

As of September 4, 2026, GWM Worlds 2 is a research preview. There is no self-serve sign-up, no waitlist tier and no public API; the launch page routes business enquiries through a contact form. Runway has not published a general availability date.

For comparison, Atlas is also closed, in early access for select partners, and its status is tracked on the Atlas API page. The state of the field is that two of the most-discussed world models announced this month are both unavailable to build on, and the World Labs model you can actually call is Marble.

What to watch next

• A developer API. The single change that would move GWM Worlds 2 from a demo to a platform. Nothing has been announced.
• Published limits on session length and memory. The sliding-window detail implies a bound; nobody has stated it.
• Independent testing of the arbitrary-actions claim. The first reviewer to push it past curated prompts will establish whether the generality is real.
• Any form of export. If Runway adds even video-plus-depth output, GWM Worlds 2 starts to overlap with the reconstruction models rather than sitting beside them.

How it differs from a video generator

This is the most common confusion about GWM Worlds 2, and it is worth settling, because Runway is best known for video generation and the two products look superficially similar.

A video generator takes a prompt and returns a finished clip. The entire output is decided before you see any of it, and if you want something different you change the prompt and generate again. The unit of work is a clip.

A world model generates as you watch, conditioned on what you do. The output is not decided in advance because part of the input has not happened yet: your next action. The unit of work is a session.

That difference has consequences beyond interactivity. A video generator can look ahead and plan a shot, so it can afford global coherence across the whole clip. A world model cannot look ahead at all, so its coherence has to come from memory of what it already produced, which is why the sliding window matters so much and why every model in this category has a horizon past which the world starts to drift.

It also changes what “good” means. A video generator is judged on the finished artefact. A world model is judged on whether the world holds together while you push against it, which is much harder to demonstrate in a trailer and much easier to break in the hands of a stranger. That gap between demo and use is the main reason to hold judgement on any world model until people outside the lab have tried it.

Bottom line

In our view GWM Worlds 2 is the most convincing interactive world model published so far, on two axes none of the models compared here covers at once: it makes sound, and it takes instructions in open language. It is also entirely unavailable, exports nothing, and has been measured by no one outside Runway.

If you are choosing a world model to build on today, this is not it, and that is a statement about availability rather than quality. Compare the alternatives at the comparison hub, or start with Atlas vs GWM Worlds 2. To hear when Atlas access widens, follow the changelog or its RSS feed.