HyperFlow H3 Flash Shrinks MiniMax-H3 From 31B to 13.7B, but Its Licence Excludes the US, EU and UK

Video Rebirth released HyperFlow H3 Flash, a 13.7-billion-parameter compressed copy of MiniMax's roughly 31-billion-parameter H3 video model, with weights whose licence excludes the United States, the EU, the UK and South Korea without MiniMax's authorisation.

The code and weights went up on GitHub and Hugging Face on Nov. 6, according to the repository's metadata. The speed and size figures below come from the project's own README and model card, so all of them are developer-supplied. The card says the project is "not an official MiniMax product."

What Video Rebirth changed in MiniMax-H3

The README says the MiniMax-H3 transformer was shrunk by structural pruning, low-rank factorisation and W8A8 INT8 quantisation. Pruning dropped 3 of 50 blocks in the fl2va variant and 1 of 50 in ref2va. A two-time conditioning scheme lets it sample in 4 to 8 steps without classifier-free guidance, with 6 steps the default.

It generates 5.2 to 15 seconds of video at 24 fps, 1344 by 768 by default, with a 32 kHz stereo soundtrack. Three modes are listed: text to video and audio, first and last keyframes (fl2va, 13.7B), and reference-based generation with up to 12 references (ref2va, 13.9B).

The card says prompt adherence, language support, biases and failure modes follow the MiniMax-H3 base model. It lists no quality benchmark. Every number in the README is a speed or a size.

Speed on a 5.2-second clip

The README gives these times for a 5.2-second clip at 6 steps in INT8.

1. 4x NVIDIA B2008.7 s
2. 1x NVIDIA H20036 s
3. 1x RTX 509092 s

Source: Video Rebirth README, accessed 2026-11-08

The fastest figure is 8.7 seconds on four B200 GPUs, which is slower than the clip's own 5.2-second playback. The README also says the model runs in real time on four Blackwell GPUs, and the card attributes that result to NVIDIA SOL-H3, an inference-acceleration partner, not to the HyperFlow package.

Consumer cards are slower, and longer clips scale worse than the length alone suggests. The README's GGUF measurements, again its own:

GPU (GGUF)	5.2-second clip	15-second clip
RTX 5090	99 s	575 s
RTX 4090	157 s	830 s
RTX 3090	315 s	1,835 s

Hardware: 24 GB of VRAM, 64 GB of RAM, a 90 GB download

The card recommends the Q4_K_M GGUF, 9.1 GB, for 24 GB GPUs, and GGUF generally for PCs with 64 GB of RAM or less. The INT8 build needs roughly 69 to 91 GiB of system RAM with offloading, and ref2va in INT8 needs 100 to 125 GB. The first-run download is about 90 GB, or about 68 GB with an FP8 text encoder.

GPUs with 16 GB have not been measured, and multi-GPU runs need Linux or WSL2. For comparison, llama-modchip's DeepSeek claim targeted the same 64 GB class of desktop for a language model.

The licence: $20 million, a US exclusion and a ban on training

The inference code is Apache-2.0, with a GPL-3.0 ComfyUI node and an MIT patch to stable-diffusion.cpp as exceptions. The weights fall under the MiniMax H3 Community License Agreement. As summarised on the card:

• The model is not licensed in the EU, UK, South Korea or the US without prior MiniMax authorisation.
• Products above $20 million in annual revenue need MiniMax's written authorisation.
• Products built on the weights must display "MiniMax H3" in their interface.
• Outputs may not be used to train or improve other AI models.
• Publicly distributed AI-generated content must be labelled as such.

The card shows about 40,800 downloads in the last month and 99 likes. The GitHub repository had 135 stars, 6 forks and a single commit. As our report on NVIDIA's speech model found, vendor-supplied figures need an independent rerun before they mean much.

Nobody outside Video Rebirth has published a quality comparison against the 31B original. The card lists four community discussions and the repository has 3 open issues, which is where any such report would appear first.