Alaya Lab releases Evoke, a 14B open world model with external scene memory for hour-long interaction
- Alaya Lab's Evoke is a 14-billion-parameter open-weights world model that stores generated scene geometry outside its video denoiser, indexed by camera pose, to maintain consistency across hour-long interactive sessions.
- After each nine-frame, 1.5-second chunk, Evoke estimates depth, converts frames to 3D geometry, and retrieves up to eight overlapping stored views to condition the next generation.
- The distilled model uses three denoising steps without classifier-free guidance and generates 1.5 seconds of 384 by 640 video in 2.11 seconds on one H200, so it remains slower than real-time playback.
- Evoke publishes its full training ladder, including a 148 GB teacher and roughly 53 GB checkpoints for several stages; the usual distilled checkpoint plus shared text encoder and VAE takes about 75 GB on disk.
- The inference weights alone occupy roughly 39 GB of VRAM before generation, while all published performance measurements use a 141 GB H200; Alaya Lab gives no minimum-memory specification.