Fast-WAM
- representation only
- 2026
- Wan 2.2 (5B)
- World action model
Model cardPaper ↗
Fast-WAM keeps video co-training while learning, then throws away future-frame generation at test time so the policy can act from latent world features in one pass.
Key facts
Rows below are the register record in lib/models.ts. A dash means the value is not confirmed from a primary source. Classification follows Reuss, NVIDIA Technical Blog (15 June 2026) and the model's own paper.
| Name | Fast-WAM |
|---|---|
| Lab | — |
| Year | 2026 |
| Paradigm | representation only |
| Backbone | Wan 2.2 (5B) |
| Source | https://arxiv.org/abs/2603.16666 |
What the paper claims
- The paper asks whether WAM gains come from video modeling during training or from imagining pixels at inference. Fast-WAM keeps the training objective and skips future prediction at test time. Yuan, Dong, Liu & Zhao, 2026
- The architecture is a Mixture-of-Transformers: a Wan2.2-5B video DiT plus a 1-billion-parameter action expert (about 6 billion parameters in total) with shared attention. At inference the video branch encodes the observation in a single forward pass. Yuan, Dong, Liu & Zhao, 2026
- Controlled variants in the paper stay close to imagine-then-execute WAMs when video co-training is kept, while removing that co-training causes a larger drop. Yuan, Dong, Liu & Zhao, 2026
What it evaluates
- The authors report competitive results on LIBERO and RoboTwin without embodied pretraining, plus a real towel-folding task, at 190 ms inference latency on a single NVIDIA RTX 5090D V2 32GB GPU. They report that this is more than 4× faster than existing imagine-then-execute WAM approaches. Yuan, Dong, Liu & Zhao, 2026
- Fig. 4 of that paper, on the same towel-folding task and GPU, labels inference latency as Fast-WAM-Joint 580 ms, Fast-WAM-IDM 810 ms, π0.5 180 ms, and Fast-WAM 190 ms. Those 580 / 810 ms figures are Fast-WAM's own variants, not a survey of WAMs in general. Yuan, Dong, Liu & Zhao, 2026
On this register
Representation only. Uses the video backbone's internal features and skips generating frames. That axis is defined in what is a world action model? and in Reuss, NVIDIA Technical Blog (15 June 2026).
Other representation only models on the register: Odyssey-3.
The same backbone string (Wan 2.2 (5B)) also appears on LingBot-VA.