Alaya-EVOKE
Released Aug 13, 2026
Open-weight, three-step, CFG-free interactive video world model from Alaya Lab with camera control and mid-rollout prompt switching, aimed at endless long-horizon rollouts.
- Output
- Live video stream
- Interactive
- Yes
- Inputs
- text, image, video, actions
- Max session
- —
- Resolution
- 384x640
- Frame rate
- 24 fps
- Access
- Generally available
- Licence
- Apache License 2.0
Strengths & weaknesses
Strengths
- Open weights (Apache-2.0) with training and inference code
- Camera-controlled rollouts with per-chunk prompt switching
- Faster than playback in the Director UI on five H200s (1.5 s chunks in 1.34 s)
Weaknesses
- Low resolution (384x640)
- Slower than real time on a single H200 (1.5 s of video every 2.11 s)
- Required depth backend is CC-BY-NC-4.0, which restricts commercial use
releaseDate = paper date on Hugging Face (arXiv 2608.13546, 'Published Aug 13'); GitHub news has a Day-0 WBench entry 2026-08-14, Evoke-Turbo (14B, controllability-tuned) 2026-09-11 and interactive UI 2026-09-12. Repo is Apache-2.0 but the required ViGeo depth model (and optional Depth-Anything-3) is CC-BY-NC-4.0, so commercialUse true applies to the repo only. 'actions' = camera pose tracks / movement and look controls. Documented coherence is 30 s rollouts; 'hour-scale' rollouts are described but no measured result (a third-party report's 65-minute figure was not found on primary pages). Predecessor AlayaWorld (2026-07-08, HF AlayaLab/AlayaWorld) not added separately. Alaya Lab's affiliation: paper lists USTC, Shanghai Innovation Institute and Alaya Lab; no lab website found.
Links
API model strings
- huggingface
AlayaLab/Evoke - huggingface
AlayaLab/Evoke-Turbo