Hao AI Lab's FastH3 Slashes MiniMax Video Generation by 14x
FastVideo distilled MiniMax's 33B video-and-audio diffusion transformer from 50 denoising steps down to four, with open weights and 90% sparse attention.
PRO- FastVideo released FastH3 Preview v1, a 4-step distillation of MiniMax H3's 33B video+audio diffusion transformer.
- Uses data-free DMD2 distillation: student trained against frozen teacher with a learned critic, no video data required.
- Adopts Video Sparse Attention at 90% sparsity with tile-64 CUDA kernel, up to 14x faster on Blackwell GPUs.
- Trained on ~258k text prompts across 32 GB200s at 768x1344, 124 frames with synchronized audio.
- Text-to-video-and-audio only; first-frame and reference-image conditioning not distilled in this preview.
- Integrated into vLLM-Omni for full-MP4 generation faster than playback; blog.
The Hao AI Lab team behind FastVideo released FastH3 Preview v1, a distilled version of MiniMax's H3 video model that cuts the denoising loop from 50 steps to 4 while preserving synchronized audio-video generation. The recommended checkpoint, FastVideo-FastH3-4-step-Preview-v1, is live on Hugging Face under the MiniMax H3 Community License, with open weights and a hosted ZeroGPU demo.
Base H3 calls its 33B audio-video diffusion transformer 49 times per generation. FastH3 reduces that to four calls and also cuts the attention work inside each call, yielding up to 14x faster text-to-video-and-audio on Blackwell GPUs according to the team's benchmarks.
Two knobs, one giant speedup
The step-count reduction uses Distribution Matching Distillation 2 (DMD2), which trains the student model with a frozen Base H3 teacher and a learned critic. The critic scores how far the student's noise predictions drift from the teacher's, and that gap becomes the training gradient. Instead of teaching the student to mimic outputs frame by frame, the method trains it to stay close to the teacher's score function in a way that works in far fewer steps.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
