More than lip sync.
See it in motion.
Every clip on this page is generated by the system. Pick a category, then turn the sound on to follow the lip sync.
4-bit attention, side by side.
Both halves come from the same 4-step model with the same speech, prompt and random seed. Only the attention precision differs. Drag the divider to compare.
16-bit precision is kept only in the final five attention blocks, the most sensitive to rounding error; all other attention runs in 4-bit (NVFP4).
Fewer steps. Faster attention.
Distillation removes most of the denoising steps; 4-bit attention speeds up the most expensive operation inside each remaining step.
Attention throughput
Show the numbers
Efficient by design.
Two complementary techniques make the model cheaper to run: one reduces how many steps it takes, the other makes each step lighter.
- Original modelMiniMax H3, 40 denoising steps
- Step distillationLearn to generate in fewer steps
- 4-step model10× fewer denoising steps
- 4-bit attentionNVFP4 with rotation and calibration
- Efficient inferenceFaster attention in every step
Distillation
Video diffusion models build each clip by refining noise over many steps. The original model uses 40. Distillation trains a version that produces the video in just 4 steps.
Quantization
Attention is one of the most expensive operations in video generation. We compute it in 4-bit floating point (NVFP4), using rotation and calibration to keep rounding errors small, so the output stays close to 16-bit attention.
From research to real-world interaction.
Efficient talking avatars open up a range of potential applications.