Agnes Video H3 MoE vs PAN 2
Compare documented capabilities and access. A dash means a comparable figure is unavailable.
| Specification | ||
|---|---|---|
| Task | Audiovisual generation | World simulation |
| Published capabilities | Experimental MiniMax H3 FL2VA adaptation that replaces selected feed-forward layers with routed sparse experts. | Predicts what happens after an action in a starting scene. |
| Access | Public composite package with original weights, learned router and adapter, and reference loader | Public weights, full training code and inference code |
| Result | Video with synchronized audio | Simulated video conditioned on a scene and action |
| Operating limits | Supports an initial frame, an ending frame or both with the publisher’s reference loader. | Not published |
| Sources | Model weights | Publisher documentation |
| Inputs | Text, Images | Text, Images |
| Weights | Open weights · MiniMax H3 Community License | Open weights · Apache-2.0 |
| Released | Nov 20, 2026 | Nov 10, 2026 |