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-generation
- reference-to-video
- text-to-audio-video
- fastvideo
- minimax-h3
- fasth3-omni
---

# FastH3 Omni — BF16 Release

**This is the final BF16 release checkpoint of FastH3 Omni.** It generates video with synchronized audio from a text prompt and ordered image, video and audio references, using eight transformer evaluations.

[All release variants](https://huggingface.co/collections/FastVideo/fasth3-omniref) · [UniServe](https://github.com/hao-ai-lab/uniserve) · [FastVideo](https://github.com/hao-ai-lab/FastVideo)

## Choose a runtime

| Hardware | Runtime and setup |
| --- | --- |
| NVIDIA GB200 / B200 | [UniServe deployment guide](https://github.com/hao-ai-lab/uniserve/blob/main/docs/minimax_h3/minimax_h3.md#deployments) |
| NVIDIA RTX PRO 6000 Blackwell | [UniServe deployment guide](https://github.com/hao-ai-lab/uniserve/blob/main/docs/minimax_h3/minimax_h3.md#deployments) |
| NVIDIA Hopper, including H100 / H200 | [UniServe requirements and setup](https://github.com/hao-ai-lab/uniserve/blob/main/docs/minimax_h3/minimax_h3.md#requirements) |
| NVIDIA DGX Spark (GB10) / RTX 5090 | [FastVideo local FastH3 Omni guide](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3_omniref.md) with the [NVFP4 QAD variant](https://huggingface.co/FastVideo/FastH3-OmniRef-NVFP4-QAD) |

For server deployment, use this BF16 checkpoint with UniServe's reference-conditioned (`ref2va`) configurations. The guide covers installation, model downloads, GPU placement, request capacities and the video/audio API. For the BF16 deployment, choose UniServe's `quality` precision preset. NVFP4 presets require Blackwell hardware.

## Local AI on DGX Spark and RTX 5090

Use **FastVideo** with the [NVFP4 QAD variant](https://huggingface.co/FastVideo/FastH3-OmniRef-NVFP4-QAD) for local inference. QAD means quantization-aware distillation; these weights are trained for NVFP4 execution on Blackwell. Start with the [FastVideo cookbook](https://haoailab.com/FastVideo/cookbook/) and [FastH3 Omni recipe builder](https://haoailab.com/FastVideo/cookbook/minimax-h3/fasth3-omniref/).

| Device | Setup and inference recipe |
| --- | --- |
| DGX Spark (GB10, 128 GB unified memory) | [Install FastVideo](https://haoailab.com/FastVideo/getting_started/installation/spark/) · [`spark` recipe](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_fasth3_omniref_spark.yaml) |
| RTX 5090 (32 GB) | [Install FastVideo](https://haoailab.com/FastVideo/getting_started/installation/gpu/) · [`rtx-5090` recipe](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_fasth3_omniref_rtx5090.yaml) |

Both profiles default to 480p. Spark keeps the model components resident; the RTX 5090 offloads the text encoder and requires a precomputed AdaLN table. Follow the [local inference guide](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3_omniref.md) to compose the model directory, prepare the table and choose a decoder. Keep the device recipe's attention and precision settings.

For FP8, see the [FastH3 Omni FP8 weights](https://huggingface.co/FastVideo/FastH3-OmniRef-FP8) and [FP8 device setup](https://github.com/hao-ai-lab/FastVideo/blob/main/docs/inference/fasth3_omniref.md#device-profiles), including the [RTX 4090 recipe](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_fasth3_omniref_rtx4090.yaml) for hardware without FP4 support. Its bundled NVFP4 text encoder requires Blackwell; non-FP4 hardware needs compatible BF16 encoder components. FP8 and NVFP4 QAD outputs can differ from BF16.

The optional [NVFP4 VAE decoders](https://huggingface.co/FastVideo/FastH3-NVFP4-VAE-Decoder) offer 36-layer quality and 8-layer fast tiers. The guide documents decoder selection with `--decoder`. These local variants have their own memory and precision settings.

For visual workflows, use the [ComfyUI variants](https://huggingface.co/FastVideo/FastH3-OmniRef-Comfy).

## Inputs and outputs

- **Inputs:** a text prompt plus up to nine images, three videos and three audio clips, with at most twelve references in total. At least one image or video is required.
- **Reference order:** use `<Picture 1>`, `<Video 1>` and `<Audio 1>` tags to refer to the ordered inputs. Video references include their soundtracks.
- **Output:** video with synchronized 32-kHz stereo audio at 24 FPS. Supported request sizes and lengths depend on the runtime and available memory.
- **Sampling:** eight transformer evaluations, guidance 1.0. The supplied configuration sets the video/audio schedule and reference-conditioned attention.

## Download

```bash
hf download FastVideo/FastVideo-FastH3-Omni-8-Step-V1 \
  --local-dir ./FastH3-Omni-BF16
```

The repository includes the BF16 reference transformer, schedulers and inference configuration. The runtime loads the shared text encoder, tokenizer, processor and video/audio VAEs from MiniMax-H3. Follow the runtime guide to prepare those components.

## 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 to [Reactor](https://www.reactor.inc/) for providing compute, and 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).