Runway unveils playable world model with sound

According to the announcement, GWM Worlds 2 is an autoregressive diffusion model that generates 720p video at 24 frames per second together with 48,000Hz audio. Sound is the biggest change from the previous version — up to now, the worlds these models produced were essentially silent films.

Runway has also effectively removed any limit on session length. The model remembers past frames through a sliding-window mechanism, meaning you can keep playing for as long as you like, according to the company.

The model draws on three streams of information: the description and initial frame used to establish the world, the current camera position and the actions a user inputs in the moment, and whatever past scenes remain inside that sliding window.

Three Seats At The Controls

The core of this announcement is that the same world can be controlled from three separate roles — much like a film set, where actors and a director each play a different part.

The player moves a character in first- or third-person view. The director doesn't talk to the character but to the scene itself, adjusting the stage by changing the weather or lighting a campfire, for instance. The agent role lets an AI move a character instead of a person, and it's built as an opening for robot simulation or agent evaluation.

Because the roles are separate, multiple people can join at once. Player 1, Player 2, and a director can each connect to the same world with their own object to control, and Runway said the demo used LiveKit, a real-time streaming technology.

World Prompt: The Name For Calling Up A World

Underpinning all this control is a format called the world prompt, which splits a world's state into two parts. One is the genesis prompt, a persistent definition of the environment and characters. The other is a timestamped stream of events.

The event side is written in free-form sentences, but each one is tagged with who it's addressed to. It's like calling out a specific actor over a radio — because the target is specified, instructions from multiple people don't get tangled together even when they overlap.

Depending on how fast prompts need to be written, there are three modes: a film-style mode where everything is prepared in advance and simply watched, a visual-novel-style mode that exchanges one turn at a time, and a real-time mode that must respond within tens of milliseconds. In real-time mode, running text through an LLM would be too slow, so pre-written sentences are bound to keys instead, according to Runway.

Even creating a world in the first place is done through language. Type something like "third-person dirt bike in a snowy landscape," and the system automatically generates the initial frame, the genesis prompt, and the control keys, Runway explained. It's also possible to feed in existing footage as the opening and then keep playing from where that footage leaves off.

Rendering Screens Versus Rendering Worlds

Placed side by side, the two announcements reveal where Runway is pushing its frame-prediction technology. One direction is software screens; the other is physical space.

Category	Solaris	GWM Worlds 2
What it renders	App and OS screens	Walkable 3D space
User input	Clicks and drags	Character control, scene direction, agent actions
Sound	Not mentioned in the announcement	Generated together at 48,000Hz
Target use	No-code interfaces, agent training	Games, video production, robot training

Both trace back to the same root: GWM-1, unveiled on December 11 of last year. At the time, Runway introduced it as a general-purpose world model that simulates reality in real time, and split it into three branches — Robotics, Avatars, and Worlds. It could produce 720p video up to two minutes long. Worlds 2 is the follow-up that grew the Worlds branch by adding sound and controls.

Still A Research Preview

This isn't a public release. The announcement page explicitly labels it a research preview and offers only an inquiry form. Pricing and a launch timeline were not disclosed in this announcement.

Runway itself lists the limitations. Real-time generation trades image quality for speed, so sweeping the camera around sharply can break up an object's texture and detail. The model also still struggles to accurately remember locations visited long ago, and no reference image can be added beyond the initial frame. Getting characters to interact with each other in complex ways still requires help from outside tools.

Editor's Take

When GWM-1 was announced last December, it seemed odd that a video company was talking about robots. Nine months later, seeing Worlds 2, the picture snaps into focus. What Runway is selling is shifting from video clips to controllable space.

Any team that has actually put video-generation tools to work will feel the weight of this shift immediately. Until now, the biggest headache with AI video has been that the only way to fix a scene you don't like is to tweak the prompt and regenerate the whole thing. With three separate control seats and the ability to address instructions to a specific target, there's now a path to editing inside a scene instead of regenerating it — changing just the weather, say, while leaving everything else untouched.

It's no coincidence that the same week saw World Labs unveil Atlas, which builds 3D spaces from a handful of photos. Add in NVIDIA expanding its physical-AI ecosystem with Cosmos 3 last month, and the battleground for the second half of this year has clearly moved from prettier single shots to spaces you can move around inside.

What teams need to prepare now isn't hardware, it's material. The world-prompt approach builds a world through text and directs its actions through text as well. That means the lore, character behavior rules, and scene directions a game studio or production company already has become the input directly. Teams that already have this documented can plug straight in the day the model opens up; teams where all of it lives only in someone's head will have to start writing it down from scratch.

A design where a player, a director, and an agent share control of one screen compresses a structure that game engines took two decades to build into a few lines of text. Runway has decided that the next stage for a video company is to become a company that sells worlds.