FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree

A dense-attention FastH3 Preview v1 ablation from FastVideo. It generates synchronized video and audio from text with four transformer forwards. This step-1000 model was trained with data-free DMD2.

Blog · Matching LoRA · Recommended checkpoint · FastH3 collection

Run with FastVideo

Install uv, then use the CUDA 13 / Blackwell path below. It selects FastVideo's published CUDA kernel wheel instead of compiling the kernel locally. See the installation guide for other platforms.

git clone https://github.com/hao-ai-lab/FastVideo.git
cd FastVideo
uv venv --python 3.12 --seed
source .venv/bin/activate
UV_TORCH_BACKEND=cu130 uv pip install \
  --no-sources-package fastvideo-kernel \
  -e ".[fasth3]"

FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN FASTVIDEO_FA4=1 \
python examples/inference/basic/basic_minimax_h3_t2v.py \
  --model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-Dense-DataFree \
  --prompt "your prompt" \
  --steps 5

Five scheduler points execute the trained four transformer forwards. The default uses four GPUs; other supported GPU counts must divide H3's 56 attention heads.

Scope

This preview supports text-to-audio-video generation. FL2VA and Ref2VA were not distilled. Difficult motion, fine detail, and some audio may remain below the base MiniMax H3 model. This checkpoint inherits the MiniMax H3 Community License.

Acknowledgements

We thank Nuva Lab for bringing production grounding to FastH3 through its experience with real-world creative video-agent workloads. Its production-aligned post-training insights help bridge open-source research to practical data-assisted distillation for commercial video workflows, with Omni Ref as the next focus.

We thank the NVIDIA FastGen team for the DMD2 framework and H3 reference experiment that helped us align the score clock, modality shifts, and backward simulation.

We also thank MiniMax for releasing H3-Base, and the vLLM project, NVIDIA, and MBZUAI for their continued sponsorship and support of FastVideo.