Models

Runway GWM Worlds 2 Streams Interactive AI Environments

Runway has unveiled GWM Worlds 2, a real-time world model that streams interactive 720p video and synchronized audio to advance how developers simulate environments and train AI agents.

AlphaSignal15 hrs agoModels
Image: AlphaSignal

Runway has introduced GWM Worlds 2, the latest iteration of its General World Model technology. The system generates interactive environments on the fly, streaming continuous 720p video at 24 frames per second alongside synchronized 48,000 Hz audio. Unlike traditional video generators that render static clips, this model processes live control signals like text actions and camera movements to continuously generate new frames. It builds on the previous GWM-1 architecture, which was based on Gen-4.5.

To manage these environments, Runway introduced a control format called WorldPrompt. This system divides a scene into a persistent genesis prompt—which defines the environment, character attributes, voices, and physical laws like gravity—and a timestamped stream of overlapping text actions. The underlying autoregressive diffusion architecture uses causal, cache-assisted video and audio decoders. It relies on a sliding KV-cache window to maintain context from past frames while evicting older data, allowing sessions to run indefinitely.

The model supports three operational modes: ahead-of-time rendering for filmmaking, turn-based generation for visual novels, and real-time streaming where keys and mouse clicks trigger instant actions. It also supports multiplayer capabilities via LiveKit, allowing users to assume roles such as player one, player two, or director. However, Runway notes that real-time generation trades fidelity for speed, leading to occasional geometric drift during rapid camera movements and imperfect long-term consistency. The model also lacks support for image references beyond the initial frame.

For practitioners, GWM Worlds 2 represents a shift in how interactive experiences and AI policies are developed. Robotics researchers can use the streaming simulator to train embodied agents in visually rich environments without deploying physical hardware. Game developers can also leverage the system to quickly turn a single-line description into a playable preset. Currently positioned alongside competitors like Google's Genie, World Labs' Marble, and NVIDIA's Cosmos, GWM Worlds 2 is available as a research preview with no public pricing, accessible only through a business contact form.

This is our own summary of reporting by AlphaSignal

More in Models