Evoke brings persistent memory to world models

Evoke’s real breakthrough is architectural: persistent world state sits outside the denoiser, making interactive video generation more practical for extended sessions. The release is promising for research, though its H200 requirements and restrictive depth-model licenses limit immediate production use.

• –Three-step, CFG-free generation improves responsiveness compared with many-step video models
• –External geometric memory helps preserve spatial continuity as users explore
• –Mid-rollout prompt changes enable controllable synthetic environments and scenario branching
• –The released stack supports text-to-video, image-to-video, video-to-video, and segmented prompt switching
• –Reported performance is strong on WBench, but hour-scale coherence still needs broader independent validation