Runway’s GWM Worlds 2 research preview can generate interactive 720p video and synchronised audio at 24 frames per second, with a new WorldPrompt system for directing characters, cameras and events. The important catch is that this is not yet a programmable game world: it can respond to prompts, but it does not expose structured state or reliable scripting.
Watch Desk analysis
What happened
Runway has introduced GWM Worlds 2 as a real-time interactive world model. WorldPrompt lets users define elements such as the opening frame, characters, environments and timestamped actions, then prompt further events while the simulation is running.
Latent Space spoke with Runway CTO Kamil Sindi, principal research scientist Robin Kahlow and co-founder and co-CEO Anastasis Germanidis. They described a system that generates continuous 720p video at 24 frames per second, alongside 48kHz audio, rather than producing a self-contained clip and stopping politely at the end.
Runway says it adapted its audio-video foundation model to the WorldPrompt format, trained it to generate content autoregressively and used distillation to make generation fast enough for real-time interaction. In plain English, the model generates the next frame or small group of frames from what has already happened, instead of rendering the whole scene in one go.
Key findings
- WorldPrompt
A prompt-based control layer for characters, cameras, environments and timestamped events. - Real-time output
GWM Worlds 2 streams 720p video at 24 frames per second with synchronised audio. - No structured state
An agent sees generated video and audio, but cannot currently read a formal game-state record from the world. - Research-preview limits
Runway says movement is relatively reliable, while harder actions and long-running interactions remain inconsistent. - Potential agent training
Runway sees simulated environments as a way to generate synthetic data and test agents at scale.
Why it matters
This is a meaningful step beyond video generation because the model is being asked to maintain an evolving scene while reacting to new instructions. That makes it relevant to games, virtual production, robot training and agent evaluation, not merely to making a striking 10-second clip.
The missing piece is control. Minecraft and Roblox work because their worlds have explicit rules, objects and state that software can inspect and update. GWM Worlds 2 currently offers an audiovisual experience without that same dependable underlying ledger. A character may move convincingly, but the system may not know in a machine-readable way whether the character has picked up a key, opened a door or merely produced a persuasive cinematic suggestion of doing so.
Our read
Runway is attacking one of the hardest problems in generative media: keeping a world responsive, coherent and fast enough to inhabit. The result looks more like an early runtime for simulated experience than a finished game engine, and that distinction matters.
The exciting possibility is an interface for creating environments without writing conventional code. The uncomfortable possibility is a world that looks consistent until a user asks it to remember something important. Prompting is a powerful control surface, but it is not the same thing as rules, causality or state management. The pixels are moving quickly; the bookkeeping is still tying its shoelaces.
What to watch
- Whether Runway exposes structured state or scripting alongside WorldPrompt.
- How long a world remains coherent before autoregressive errors accumulate.
- Whether actions produce consistent, plausible consequences rather than attractive improvisations.
- Whether robotics and agent-testing partners use the system outside demonstrations.
Discussion spark: Should real-time world models first be judged as creative interfaces, or should they be held to game-engine standards of state, rules and causal reliability?
Sources and evidence
- Runway’s WorldPrompt and the Engineering of Real-Time Worlds (25 September 2026, 01:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.