GWM-1

GWM-1 (General World Model 1) is Runway's first general world model family, announced on December 11, 2025. It is an autoregressive model built on the Gen-4.5 foundation, generating video frame by frame in real time and controllable interactively via action signals: camera pose, robot commands, or audio. Runway positions world models as the frontier of AI progress — environments where agents can explore, make mistakes, and learn via trial-and-error in simulation, far faster than in the real world.

The model comes in three post-trained variants:

GWM Worlds — a real-time environment simulation model. Given a static scene, it generates an immersive, infinite, explorable space as the user moves through it, with geometry, lighting, and physics. Key feature: spatial consistency — turning in place preserves the background, walking forward and back returns to the same point. Use cases: gaming, education, VR and immersive experiences, training sandboxes for AI agents.

Runway Characters (GWM Avatars) — an audio-driven interactive video generation model that simulates natural motion and expression for arbitrary photorealistic or stylised characters. Renders realistic facial expressions, eye movements, lip-sync, and gestures during both speaking and listening, sustaining quality across long conversations. Available in Runway API and the web product. Use cases: real-time tutoring, customer support, training simulations, interactive entertainment.

GWM Robotics — a robotics-focused model using a video-generation diffusion backbone (trained at the same scale and pretraining as Runway's frontier creative models). Three use cases: (1) policy execution — diffusion backbone for robot policies, improving generalisation and sample efficiency; (2) synthetic data augmentation — generating training data for robotics datasets (novel objects, instructions, environmental variations); (3) policy evaluation in simulation — testing policies (e.g. VLAs like OpenVLA, OpenPi) in simulation instead of on physical robots. Available via the GWM-1 Robotics SDK (Python).

Key specifications: the model generates video up to 2 minutes long in 720p, conditioned by actions (camera, events, robot pose, speech), with fine-tuning on custom action domains. A customisable model is available — Runway provides a developer portal for fine-tuning (dev.runwayml.com).

Wider context: Runway announced the General World Models research direction two years before GWM-1 (2023). The company believes language models alone won't solve the world's hardest problems — robotics, disease, scientific discovery — because real progress requires models that experience the world and learn from their own mistakes. The next step after GWM-1 is unifying many domains and action spaces under a single base world model. Announcement: runwayml.com/research/introducing-runway-gwm-1.