Runway GWM-1

What it is

GWM-1 is the model Runway announced on 11 December 2025 as its first "General World Model" ^[1]. Runway defines a world model as an AI system that builds an internal representation of an environment and uses it to simulate future events in that environment ^[3]. Applied to GWM-1, that means the system produces video of a scene that keeps moving forward in time, and the direction it takes is driven by input rather than fixed when generation begins.

The distinction matters because a video generator and a world model are asked to do different things. A video generator is judged mostly on how the frames look. A world model is judged on whether the environment holds together while it is acted on: if the viewpoint turns away and then turns back, what was behind it should still be there, in the same place. Holding that consistency is what makes the output usable as a simulated place rather than as a clip.

How it works

GWM-1 is autoregressive: it generates frame by frame, in real time, instead of producing a whole clip in a single pass ^[1]. In an autoregressive model, each step's output becomes part of the condition for the next step, so frames that already exist are used to decide what the next frame looks like. Because that loop keeps running while the video plays, the output can react to input during generation rather than being locked in at the start. Real-time generation is what allows that reaction to be visible as control rather than as a correction applied afterwards.

What the model is conditioned on is an action. GWM-1 conditions on camera pose, robot commands and audio ^[3]. Camera pose moves the viewpoint through the scene, robot commands steer a simulated manipulator, and audio drives what happens on screen; each of these corresponds to one of the three variants described below. Output is 720p video and can run for up to 2 minutes ^[1].

GWM-1 is built on top of Gen-4.5 ^[1].

Three variants

At launch GWM-1 came in three forms ^[1]. GWM Worlds provides explorable environments, in which the user moves through a scene. GWM Avatars provides conversational characters and is sold under the name Runway Characters. GWM Robotics is aimed at robotic manipulation.

The three were separate post-trained models at launch, and Runway said it was working toward unifying the different domains and action spaces under a single base model ^[3]. Post-training is the stage that follows the initial training of a base model: the base model picks up general structure from large amounts of data, and post-training adapts it to a particular task or a particular kind of control signal. Keeping three post-trained variants means each can be tuned to its own action space — camera paths, dialogue, manipulation — while sharing a common foundation. Unifying them would mean one model that accepts all three kinds of action without being split into separate versions.

How it is available

Runway Characters is available through the Runway API and the web product. GWM Robotics is offered by request, together with a Robotics SDK ^[1]. The split reflects how the variants are used: a character model can be called like any other model in a product, while a manipulation model has to be attached to a robot and its control stack.

Where it sits

In February 2026 Runway raised a $315 million round, which it said was raised to pre-train next-generation world models ^[2]. Pre-training is the first and largest stage of building a model of this kind — the stage at which the model is exposed to enough material to pick up the regularities of a domain before any task-specific tuning. Describing the round that way places GWM-1 as a first generation: the variants exist and can be used, while the unified model the company describes as the next step is still to be trained.

That also sets the open question for the line. Whether one base model can absorb camera control, robot commands and audio at once, instead of three post-trained systems sitting on a shared foundation, is the problem Runway says it is working on ^[3].