Evoke

World models usually forget what is behind the camera; Evoke tackles that with an external world state bank that stores scene geometry indexed by camera pose, letting rollouts revisit earlier viewpoints without drift. The AlayaLab model is a 14B autoregressive diffusion transformer over latent chunks with multi-term memory carrying recent and distant context, distilled down to 3-step CFG-free sampling so each step costs one forward pass instead of two, and it leads the WBench benchmark against 50-step baselines. Users steer the camera and swap prompts mid-generation across text-to-video, image-to-video, and video-to-video modes. On a single H200 it produces 1.5 seconds of 384x640, 24fps video every 2.11 seconds, approaching real-time interaction. The August 2026 release ships training and inference code, four staged checkpoints plus teacher weights, and example data under Apache-2.0, though some vendored depth components carry CC-BY-NC terms. Setup is research-grade Python with datacenter GPU expectations.