MWA-Phoenix
Anyverse Dynamics open-sourced MWA-Phoenix, a 7B latent world action model for robot manipulation that uses only RGB input and runs on the robot.
The model uses DINOv3 as its vision backbone and distills from VGGT for geometry and V-JEPA for temporal change. A causal semantic VAE compresses the features, and a video expert and an action expert are linked through joint attention so it predicts scene change and robot action together. Two companion releases, Phoenix-ILA and Phoenix-PhysCanvas, target the gap between simulation and real hardware.
- Date
- Thursday 22 October 2026
- Lab
- Anyverse Dynamics
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| LIBERO-Plus | 94.63% | company |
| LIBERO-Pro | 66.48% | company |
| RoboCasa GR1 Tabletop | 82.67% up 7.5 points from the previous WALA model | company |
| RoboTwin 2.0 | 72.42% dual-arm benchmark | company |
| Bench2Dex average | 45.6% 17.3 points ahead of the strongest baseline, pi0.5 | company |
| Parameters | 7B | company |
All benchmark figures are the company's own. The license was not stated on the page, and code is on GitHub with a technical report.
Sources
This record was checked against its sources on 22 October 2026. How we check