README.md

---
license: other
license_name: minimax-h3-community-license-agreement
license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
base_model: MiniMaxAI/MiniMax-H3
tags:
- video
- vae
- decoder
- nvfp4
- fp4
- blackwell
- fastvideo
- minimax-h3
---

# FastH3 NVFP4 VAE Decoder

Blackwell NVFP4 video decoders for MiniMax-H3 and FastH3, including FastH3 Omni. They replace the video decoder while retaining the reference encoder, audio VAE and video latent space.

[FastVideo](https://github.com/hao-ai-lab/FastVideo) · [FastH3 Omni decoder setup](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3_omniref.md) · [FastH3 Omni release collection](https://huggingface.co/collections/FastVideo/fasth3-omniref)

## Choose a decoder

| Tier | Folder | Decoder layers | Resident decoder weights |
| --- | --- | ---: | ---: |
| Quality | `quality-36L/` | 36 | 1.29 GiB |
| Fast | `fast-8L/` | 8 | 0.31 GiB |

Choose the quality tier for closer reconstruction or the fast tier for lower decode cost with a quality tradeoff. Both tiers are lossy.

## Use with FastVideo

Download a packed decoder:

```bash
hf download FastVideo/FastH3-NVFP4-VAE-Decoder \
  quality-36L/decoder-nvfp4.safetensors --local-dir ./fasth3-vae
```

Follow the [FastVideo FastH3 Omni guide](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3_omniref.md) and pass the downloaded file with `--decoder` to `examples/inference/basic/basic_fasth3_omniref_pdd.py`. Use `fast-8L/decoder-nvfp4.safetensors` to select the fast tier instead.

The 36-layer decoder needs a full 36-layer MiniMax-H3 host VAE. The 8-layer tier also works with the light H3 VAE. The FastH3 Omni guide explains model composition and device profiles.

For UniServe deployment of the complete FastH3 Omni model on GB200/B200, RTX PRO 6000 Blackwell or Hopper, follow the [UniServe guide](https://github.com/hao-ai-lab/uniserve/blob/main/docs/minimax_h3/minimax_h3.md) with the [BF16 model](https://huggingface.co/FastVideo/FastVideo-FastH3-Omni-8-Step-V1). These files are FastVideo decoder components, not complete diffusion checkpoints.

## Hardware

The FP4 decoder kernels require NVIDIA Blackwell: GB200/B200, RTX PRO 6000 Blackwell, RTX 50-series or DGX Spark (GB10). Follow the [local FastH3 Omni guide](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3_omniref.md#device-profiles) for Spark and RTX 5090 setup. Hopper and older GPUs should use the original H3 video decoder.

## Files

Each tier includes:

- `decoder-nvfp4.safetensors`: packed NVFP4 weights for inference.
- `decoder.safetensors`: BF16 master weights.
- `config.json`: decoder architecture and quantization configuration.

## Quality and speed

Reconstruction against the original FP32 decoder on the same latents; lower LPIPS and higher PSNR are better:

| Evaluation clips | Quality tier LPIPS / PSNR | Fast tier LPIPS / PSNR |
| --- | ---: | ---: |
| FastH3 V2 generations, 9 clips at 480p | 0.0104 / 37.9 dB | 0.0684 / 28.5 dB |
| FastH3 Omni generations, 20 clips | 0.0039 / 46.0 dB | 0.0276 / 36.1 dB |
| Real videos, VAE-encoded, 6 held-out clips | 0.0038 / 45.0 dB | 0.0330 / 34.3 dB |

Warm video-decode time for one 832×480, 124-frame clip on one GB200, using FastVideo's fused NVFP4 kernels and batching all spatial tiles in one decoder call:

| Decoder | Seconds per clip |
| --- | ---: |
| Original FP16 decoder with `torch.compile` | 0.98 |
| Quality, 36 layers | 0.53 |
| Fast, 8 layers | 0.13 |

These are decoder-only measurements. Full video/audio generation also includes encoding, denoising and media output.

## Local AI resources

Run FastH3 Omni on **DGX Spark or RTX 5090 through FastVideo** using the [FastVideo-format NVFP4 QAD weights](https://huggingface.co/FastVideo/FastH3-OmniRef-NVFP4-QAD). Start with the [FastVideo cookbook](https://haoailab.com/FastVideo/cookbook/) and [FastH3 Omni recipe builder](https://haoailab.com/FastVideo/cookbook/minimax-h3/fasth3-omniref/); use the [Spark recipe](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_fasth3_omniref_spark.yaml) or [RTX 5090 recipe](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_fasth3_omniref_rtx5090.yaml) and the [local inference guide](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3_omniref.md) for the AdaLN table, offload and decoder setup.

[FP8 weights](https://huggingface.co/FastVideo/FastH3-OmniRef-FP8) are a separate FastVideo variant; the [FP8 recipe for RTX 4090](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_fasth3_omniref_rtx4090.yaml) illustrates the non-FP4 path with layerwise offload and compatible BF16 encoder components. Use each runtime's matching checkpoint format.

## Acknowledgements

FastH3 Omni was built in close collaboration with [Nuva Lab](https://nuvalab.ai/fasth3/), which co-developed the model with FastVideo, defined and curated the omni reference dataset, and brought production-grounded creative video workflows to FastH3. For dedicated production deployment, fine-tuning on your own data and a creative agent on top, see [Nuva Lab's FastH3-Omni page](https://nuvalab.ai/fasth3/).

Thanks also to the NVIDIA FastGen team.

## License

Use of these weights is governed by the [MiniMax-H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE).