FastVideo-FastH3-4-step-Preview-v1-LoRA

The compact LoRA releases for FastH3 Preview v1 from FastVideo. Each adapter reconstructs one four-forward FastH3 transformer from the MiniMax H3 base model.

The three VSA adapters require FastVideo's VSA-H3 backend and kernel. Use FastVideo's launchers rather than a generic PEFT loader; these adapters also contain exact delta and VSA gate tensors.

Run the recommended adapter

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]"

bash examples/inference/basic/run_fasth3_lora_preview_vsa_datafree.sh \
  --prompt "your prompt" \
  --no-warmup \
  --repeats 1

Set FASTH3_LORA_STRENGTH to change the adapter strength from its default of 1.0. The tested defaults use four B200 GPUs. On other multi-GPU CUDA systems, follow the installation guide and add --no-replicated-dit --vsa-kernel triton --no-fa4 to a VSA launcher. The GPU count must divide H3's 56 attention heads.

Variants

Adapter	Full checkpoint	FastVideo launcher
VSA / Data-Free	VSA / Data-Free	`run_fasth3_lora_preview_vsa_datafree.sh`
VSA / Synthetic, step 1300	VSA / Synthetic, step 1300	`run_fasth3_lora_preview_vsa_synthetic_step1300.sh`
VSA / Synthetic, step 1900	VSA / Synthetic, step 1900	`run_fasth3_lora_preview_vsa_synthetic_step1900.sh`
Dense / Data-Free	Dense / Data-Free	`run_fasth3_lora_preview_dense_datafree.sh`

The VSA launchers select VSA-H3 automatically; the dense launcher disables it. This preview is for text-to-audio-video generation and 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.