WAM-H3
A single-stream action model on a 33B omni-transformer
2026-09
Built a single-stream action model on MiniMax-H3, a 33B omni-transformer: robot actions take the audio slot and share one attention stream with the instruction, observation and future video, with no separate action expert. An asymmetric mask lets one KV-cache pass serve action-only denoising: ten Euler steps give a 32-step action chunk without denoising video. Trained with LoRA on LIBERO to a development checkpoint.
- Backbone
- MiniMax-H3, 33B
- Sequence
- 391 rows: text 64, observation 98, proprioception 1, actions 32, future video 196
- Trainable
- LoRA, 157M parameters at rank 16
- Inference
- One prefill pass, then ten action-only Euler steps