FastVideo · Hao AI Lab  FastH3 Preview v1

Text → video + audio
in four steps.

A 35B text-to-audio-video model that generates synchronized video and sound from a prompt in just four transformer forwards — distilled data-free with DMD2 and VSA-H3 sparse attention.

35B parameters 4 forward passes BF16 90% VSA sparsity base · MiniMax H3
🤗 Model card 📝 Blog & samples ⚙️ FastVideo repo

What it is

FastH3 is the recommended FastH3 Preview v1 checkpoint from FastVideo. Give it a text prompt and it returns a short clip with matching, synchronized audio — the picture and the sound are generated together in one pass.

What makes it notable is the speed: full-quality diffusion video models typically take dozens of denoising steps. This checkpoint (step 1300) was distilled so the whole generation runs in four transformer forwards, using a data-free DMD2 objective and the VSA-H3 attention backend at 90% sparsity — while staying close to the base MiniMax H3 model.

At a glance

4×
transformer forwards / generation
35B
parameters · BF16 weights
90%
VSA-H3 attention sparsity
56
attention heads (must divide GPU count)

How the 4-step pipeline works

Input
Text prompt
A description of the scene and its sound.
→
Backbone
FastH3 DiT + VSA-H3
Distilled transformer with 90%-sparse attention kernels.
→
Decode
Video + Audio VAE
Joint decode into synchronized frames and waveform.
1forward
2forward
3forward
4forward

The four-forward schedule is the trained default. Distillation uses data-free DMD2 (Distribution Matching Distillation) so no paired training videos are required, and VSA-H3 keeps attention at 90% sparsity to fit the compute budget.

Run it on your own hardware

The CUDA 13 / Blackwell path below pulls FastVideo's published CUDA kernel wheel instead of compiling locally. See the installation guide for other platforms.

1 · Install

# requires uv → https://docs.astral.sh/uv/getting-started/installation/
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]"

2 · Generate

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
Hardware. The tested defaults use four B200 GPUs and the trained four-forward schedule. On other multi-GPU CUDA systems, follow the installation guide and add --no-replicated-dit --vsa-kernel triton --no-fa4. The GPU count must evenly divide H3's 56 attention heads.

Prefer only the distilled adapter? Grab the matching VSA-datafree LoRA instead of the full checkpoint.

Scope & limitations