Runway GWM Worlds 2 Continuously Generates Interactive 720p Video with a Sliding KV Cache

Runway unveiled GWM Worlds 2 on September 3, transforming a video model that originally generated fixed-length clips into an interactive simulator that can continuously accept input. The official demos show output at 720p and 24 frames per second, with 48 kHz audio. Users can control characters and the camera independently, use text to instruct specific characters to move, speak, or manipulate objects, and even change the weather or the entire scene directly.

The core interface, WorldPrompt, divides state into two layers. The persistent layer contains the initial frame, scene configuration, character attributes, and rules such as gravity and collision behavior. The event layer contains actions, dialogue, sounds, and frame-by-frame camera movements, each with defined start and end times. Compared with packing every instruction into a single prompt, this representation more closely resembles a game engine’s world state and event queue. It also allows different users or agents to control individual characters and director-level events independently.

The model itself uses an autoregressive diffusion architecture. At each step, it is conditioned simultaneously on the global prompt, control signals for the current frame, and existing audiovisual tokens within a sliding window. Older frames are evicted from the KV cache, while the audiovisual decoder also uses causal caching to reduce latency. This explains how the system can generate without determining the clip length in advance, but it also exposes the limits of its long-term memory: once information leaves the cache window, geometry, materials, and character details must be reconstructed by the model rather than preserved in an explicit state database.

For engineering teams, a notable feature is that the same interface supports three execution modes: pre-scripted, turn-based, and real-time. It could therefore serve as a shared backend for game prototyping, embodied-agent testing, or generative interfaces. However, Runway acknowledges that pre-scripted operation currently delivers higher quality, while real-time mode largely maps prewritten text actions to keyboard and mouse inputs. Rapid camera rotation can cause texture and geometry drift, and reference imagery is limited to the first frame or prefilled video. Runway has not released model weights, a research paper, latency distributions, hardware requirements, or independent benchmarks for physical consistency. As a result, “real-time” should not be treated as synonymous with a deployable low-latency simulator. The next questions to watch are how well the system preserves state over long durations, its actual action-to-frame latency, and whether agents can use these worlds to complete tasks in reproducible evaluations.