Open weights · MiniMax-H3 distilled · Data-Free DMD2

Text → video + audio
in four steps.

FastVideo FastH3 Preview v1 generates synchronized video and stereo audio from a single prompt using just four transformer forwards — up to 14× faster than Base H3 on NVIDIA Blackwell, with VSA sparse attention at 90% sparsity.

4
transformer forwards
14.4×
speedup @ 15s / 1×B200
90%
attention sparsity (VSA)
T2VA
video + synced audio
Benchmarks

Up to 14.4× faster than Base H3

End-to-end latency in seconds on NVIDIA B200. Preview v1 uses four DiT calls with VSA sparse attention; speedup is shown over dense Base H3 on 1× / 4× GPUs.

Model / runtimeDuration1× B200 (s)4× B200 (s)8× B200 (s)Speedup (1× / 4×)
Base H3 · Dense FA45 s132.540.61.0× / 1.0×
10 s377.4108.71.0× / 1.0×
15 s678.7193.11.0× / 1.0×
Preview v1 · VSA Data-Free (90% sparse)5 s16.26.16.848.16× / 6.65×
10 s31.112.011.6612.13× / 9.03×
15 s47.215.512.8814.38× / 12.48×
Preview v1 · Dense Data-Free (FA4)5 s18.36.87.24× / 5.97×
10 s50.215.07.52× / 7.25×
15 s91.325.67.43× / 7.54×
FastH3 can generate 15s of 768p video in under 13s — sub-realtime — on 8×B200. Numbers from the Hao AI Lab release blog.
Run it yourself

Reproduce these on your own GPUs

The recommended checkpoint requires FastVideo's VSA-H3 attention backend. It targets multi-GPU Blackwell (the tested defaults use four B200s); the GPU count must divide H3's 56 attention heads.

 install.sh
# CUDA 13 / Blackwell path (uses prebuilt kernel wheel)
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]"
 generate.sh
python examples/inference/basic/basic_fasth3.py \
  --model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree \
  --prompt "your prompt" \
  --no-warmup \
  --repeats 1

# non-B200 multi-GPU CUDA, add:
#   --no-replicated-dit --vsa-kernel triton --no-fa4
Why no live generation here? This checkpoint needs multiple Blackwell GPUs and FastVideo's custom VSA-H3 CUDA kernel — beyond what a hosted Space provides. This Space is a curated showcase; use the commands above to run it on your own hardware, or grab the matching LoRA for just the distilled adapter.