Runway's GWM Worlds 2 research preview can generate an interactive world as a continuous 720p video stream at 24 frames per second, with synchronized audio. Its new WorldPrompt format lets a creator anchor parts of the scene, including the opening frame, and describe actions that should occur at specific times.
WorldPrompt acts as a control layer for characters, cameras and environments. Actions can also be added while the simulation is running, allowing an on-screen character to approach, speak or move in response to a new instruction. Runway trained its audio-video model on the format, then adapted it to generate a few frames at a time rather than an entire clip and used distillation to reduce the work needed for real-time output.
The system is not a game engine. WorldPrompt is natural-language prompting rather than a programming language, so it cannot guarantee state or enforce scripted rules like a conventional Minecraft or Roblox world. Runway says movement follows instructions relatively reliably, but more difficult actions remain inconsistent.
Long sessions also expose an unresolved technical limit. Generated frames are fed back into the model, so small errors accumulate, while retaining too much past context consumes GPU memory. The model does not have perfect long-term memory. Those constraints make Worlds 2 a useful research demonstration of live generative environments, not yet a replacement for deterministic game logic.