# FastH3-Omni: open weights for MiniMax H3, with sound and references

> On October 20, 2026, Nuva Lab and FastVideo released the weights for FastH3-Omni, a faster version of the MiniMax H3 video model. A 5–15 second clip comes out with stereo sound in 8 transformer evaluations instead of 50. On eight GB200 GPUs, the authors measured end-to-end latency 5.6–11.9× lower than the base model.

Published 2026-10-26 · Updated 2026-10-26
URL: https://ai2face.com/news/fasth3-omni

On October 20, 2026, Nuva Lab, together with FastVideo from Hao AI Lab, released the weights for **FastH3-Omni**. It's a distillation of [Hailuo](/models/hailuo) [MiniMax H3](/models/minimax-h3): the same set of image, video, and audio references, and a clip with stereo sound in one pass, but the transformer runs 8 evaluations instead of 50. Nuva Lab compares that input set to [Seedance](/models/seedance) 2.5. Nuva's closed production setup uses separate data and separate hosting. It isn't part of the released weights.

## What FastH3-Omni can do

According to Nuva Lab's FAQ, the open model accepts:

- text;
- up to 9 photos, and up to 3 in realtime mode;
- up to 3 video clips of 2–15 seconds each, no more than 15 seconds in total;
- up to 3 audio clips with voice or music, also 2–15 seconds each, and only together with a photo or a video;
- 12 files at most;
- a first and last frame in the same request, plus a continuation of a clip that's already finished.

The output is a clip of **5–15 seconds at 24 frames per second, with sound**. On the base MiniMax H3 this is built from, the sound is 32 kHz stereo, and stable speech is claimed for 11 languages, including Russian. Resolution up to 2K on the base model is a separate regeneration step. In the speed measurement, FastH3-Omni's target frame is 1344×768.

## How much faster FastH3-Omni is than MiniMax H3

The numbers below are FastVideo's measurement, not an independent test. They compared full serving recipes on eight NVIDIA GB200s: reference encoding, generation, the decoder, and the finished mp4. 27 matched tasks, three clip lengths, frame 1344×768.

| | FastH3-Omni | MiniMax H3, base |
|---|---|---|
| Transformer evaluations | 8 | 50 |
| End-to-end latency | 5.6–11.9× lower | baseline |
| Test frame | 1344×768 | the same |
| Realtime | on particular UniServe tasks, 768p, 5 / 10 / 15 s | not in this measurement |

“Real time” here doesn't mean every clip. FastVideo uses that wording only for particular tasks on its UniServe server.

## How to run MiniMax H3 locally with FastH3-Omni

The weights and the training code are in the FastVideo collection on Hugging Face. The technical write-up is on the Hao AI Lab blog. FastVideo describes several local options:

| Hardware | Memory | How it runs |
|---|---|---|
| NVIDIA DGX Spark (GB10) | 128 GB unified memory | `spark` recipe; every model component stays in memory |
| GeForce RTX 5090 | 32 GB | the text encoder is offloaded; a precomputed AdaLN table is required |
| GeForce RTX 4090 | 24 GB | a recipe with layerwise offload |
| ComfyUI | — | a separate FP8 checkpoint of FastH3-Omni |

The DGX Spark and RTX 5090 profiles use **NVFP4** weights — trained for compute in that format on Blackwell GPUs — and they default to **480p**. The speed from eight GB200s doesn't carry over to these machines. Nuva Lab also mentions running it on Apple Silicon through MLX.

MiniMax H3 is released under the MiniMax H3 Community License. That's MiniMax's community license, not Apache and not MIT. The commercial terms are in the agreement on the weights page, not in the announcement of the faster version.

## How to try the video in AI2Face

MiniMax H3, the model FastH3-Omni is built on, is available in the AI2Face generator as [Hailuo](/models/hailuo). Next to it is [Seedance](/models/seedance), the model Nuva Lab compares the reference set with. You pay by card or with cryptocurrency, and the price is shown before you generate. A comparison of H3 with Hailuo 2.3, prices, and clip examples are on the [Hailuo and MiniMax H3](/models/hailuo) page. For an overview of video models, see our guide to [AI video generators](/blog/nejroseti-dlya-video).

## FAQ

### What is FastH3-Omni?

Open weights for a faster MiniMax H3. Nuva Lab released them on October 20, 2026, together with the FastVideo lab (Hao AI Lab, University of California, San Diego).

### How much faster is FastH3-Omni than MiniMax H3?

In FastVideo's measurement, end-to-end latency is 5.6–11.9× lower on eight NVIDIA GB200s, across 27 matched tasks, at 1344×768. The base model uses 50 transformer evaluations; FastH3-Omni uses 8. There's no independent check.

### Can I run FastH3-Omni on my own computer?

Yes. The weights and the code are public. FastVideo has ready-made profiles for an RTX 5090 with 32 GB and a DGX Spark with 128 GB of memory. By default they render 480p, not the 768p from the GB200 benchmark. For the RTX 4090, a recipe with layerwise offload is described.

### How do I run MiniMax H3 in ComfyUI?

Through the FP8 checkpoint of FastH3-Omni for ComfyUI, from the FastVideo collection on Hugging Face. The NVFP4 weights are meant for NVIDIA Blackwell GPUs.

### How can I try MiniMax H3 without my own GPU?

In AI2Face: MiniMax H3 is available in the generator as Hailuo. You pay by card or with cryptocurrency, and the clip price is shown before you generate.

### What license does FastH3-Omni use?

The base MiniMax H3 is released under the MiniMax H3 Community License. That isn't a standard open license. Read the agreement on the weights page before commercial use.

## Sources

- https://nuvalab.ai/fasth3/
- https://haoailab.com/blogs/fastvideo-omniref-update/
- https://www.minimax.io/news/minimax-h3-open-source