Hugging Face has introduced Evoke, an interactive world model designed to overcome the limitations of existing systems in persistent memory, responsive interaction, and long-horizon generation. Traditional models face a trade-off between session length and retained memory, as maintaining history in the denoiser context or key-value cache leads to growing computational costs. Evoke addresses this by externalizing the world state into a camera-indexed bank, retrieving only view-relevant information to keep the denoiser context bounded.
The model also features a redesigned teacher for long-horizon supervision. Instead of a fixed generator, the teacher uses sparse attention with chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state. This design achieves linear growth in memory and compute while enabling supervision over long horizons, exposing content drift that remains locally plausible within short windows.
Evoke employs a 30-second distribution-matching objective under self-forced rollouts, transferring capabilities to a three-step student that uses no classifier-free guidance. This improves resistance to long-term drift while preserving responsive conditioning. The result is a model that supports open-ended, continuously evolving generation with bounded context and recurrent external memory.
On a single H200 GPU at 384×640 resolution, Evoke generates each 1.5-second chunk in 2.11 seconds. It achieves state-of-the-art performance on WBench while remaining competitive on VBench-Long and VBench-2.0. The code and demo are available on GitHub and the project page.