Runway launches GWM-1 general world model
Runway has introduced GWM-1, its first family of real-time, action-controllable world models, with separate versions for explorable environments, conversational avatars and robotics, putting the video lab directly into the world-model race.
Runway announced GWM-1, which it calls a "General World Model" built to simulate reality in real time. The model is autoregressive and built on Runway's Gen-4.5 video model. It generates video frame by frame and responds to actions such as camera pose, robot commands and audio.
GWM-1 comes in three versions. GWM Worlds produces explorable environments for gaming, education, VR and agent training. GWM Avatars produces audio-driven conversational characters with facial expressions, eye movement, lip-sync and gestures. Runway says the avatars can hold extended conversations without losing quality. GWM Robotics is a video diffusion backbone for robot policies, and Runway says it is pretrained at the same scale as its creative models.
These are currently separate post-trained models. Runway says it is working toward a single base model that covers many domains and kinds of action.
For robotics, Runway lists three uses: as a backbone for running policies, as a source of synthetic training data, and as a simulator for evaluating policies. The simulator would let developers test vision-language-action models such as OpenVLA or OpenPi without physical hardware. It predicts video conditioned on robot actions and can generate counterfactual trajectories. A Python SDK supports multi-view video and long-context sequences.
Runway also added native audio generation, audio editing and multi-shot video editing to Gen-4.5. Multi-shot editing carries a change made in one scene through an entire video. These features are "coming soon" to the web product.
GWM-1 is available only through an early-access request form. Runway calls the model "state-of-the-art" but has not published benchmarks, pricing or technical details.