HyperFlow H3 Flash

Video Rebirth's 14B compressed MiniMax H3: video with sound in 4 to 8 steps, a 9.1 GB GGUF for 24 GB GPUs, with H3's territory limits.

In short. HyperFlow H3 Flash is a smaller, faster version of the MiniMax H3 video model, made by the Singapore start-up Video Rebirth. It makes a clip of 5 to 15 seconds with matching stereo sound in 4 to 8 steps, and its 9.1 GB GGUF file runs on a 24 GB graphics card. It keeps H3's license, which does not cover the United States, the European Union, the United Kingdom or South Korea.

Summary

HyperFlow H3 Flash is Video Rebirth's compressed, few-step version of MiniMax H3, and it first appeared on Hugging Face on 4 November 2026. Video Rebirth is a Singapore start-up founded in 2024, and the card says this is not an official MiniMax product. It keeps H3's job of generating video and synchronized 32 kHz stereo audio together, in clips of 5.2 to 15 seconds at 24 frames per second, 1344×768 by default. Input can be text, a first frame or first and last frames, or up to 12 references (9 images, 3 videos and 3 audio clips, though GGUF and ComfyUI runs take images only). Sampling takes 4, 6 (the default) or 8 steps and uses no classifier-free guidance.

To shrink the model, the card says Video Rebirth pruned 3 of the 50 transformer blocks in the frame model and 1 in the reference model, factored most linear layers into low-rank pieces, quantized 180 linear layers to W8A8 INT8 behind a Hadamard rotation, and trained with two-time conditioning that allows few-step sampling. That leaves 13.7 billion parameters for the frame model and 13.9 billion for the reference model, less than half of H3's transformer. The card gives speed and memory results but no quality scores against the full H3, so judge the output yourself. Video Rebirth measures 8.7 seconds for a 5.2-second clip on four B200 GPUs, 36 seconds on one H200 and 92 seconds on an RTX 5090, and NVIDIA's SOL-H3 engine runs it in real time on four Blackwell GPUs. In ComfyUI, a 24 GB RTX 4090 takes 152 seconds for a 5.2-second clip with INT8, and a 15-second clip takes about 12 minutes. These timings are Video Rebirth's own.

The transformer comes as an INT8 file of about 16 GB or as GGUF files from Q3_K_M (7.2 GB) to Q8_0 (14.9 GB), with Q4_K_M at 9.1 GB recommended. You still need H3's companion files: the text encoder, which is 15.7 GB in the NVFP4 build and up to 51.5 GB in BF16, and the video and audio autoencoders, and the Python package downloads about 90 GB on first run. Video Rebirth measured the GGUF route on 24 GB and 32 GB RTX GPUs in PCs with 16 GB of system RAM, with peak GPU memory of 22 to 28 GiB, and says 16 GB GPUs are not measured yet. The weights are a Model Derivative of MiniMax H3 under the MiniMax H3 Community License, so its limits apply: the EU, UK, South Korea and US are excluded without MiniMax's authorization, products earning over US$20 million a year need written authorization, products must display "MiniMax H3", public clips must be labeled as AI-generated, and outputs may not be used to train other models. The inference code is Apache 2.0, the ComfyUI node GPL-3.0 and the stable-diffusion.cpp patch MIT. In the 30 days to 11 November 2026, Hugging Face counted about 242,000 downloads.

Good for

  • Making short clips with sound on one 24 GB or 32 GB RTX GPU, if you are inside the license's territory
  • Faster iteration than the full H3, with 4, 6 or 8 sampling steps and ComfyUI, Unsloth Studio and stable-diffusion.cpp support
  • Reference-guided clips with up to 12 references through the Python package, which takes video and audio references
  • Data-center serving that aims for real time, using NVIDIA's SOL-H3 engine on four Blackwell GPUs

Watch out for

  • The license is H3's, which excludes the European Union, the United Kingdom, South Korea and the United States and sets revenue, attribution and AI-disclosure conditions
  • It is not an official MiniMax release, and the card gives no quality scores against the full H3
  • You also need H3's text encoder and autoencoders, so a first download is about 90 GB through the Python package
  • GGUF and ComfyUI accept image references only, and the Python package asks for 69 GiB or more of system RAM when it offloads
  • All speeds are Video Rebirth's own, and 16 GB GPUs are not measured yet

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of HyperFlow H3 Flash. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
FL2VA INT8 (text and endpoint frames) INT8 13.7B 16 GB 18 GB Gaming GPU
FL2VA GGUF Q4_K_M INT4 13.7B 9.1 GB 24 GB Gaming GPU
FL2VA GGUF Q8_0 INT8 13.7B 15 GB 24 GB Gaming GPU
Ref2VA INT8 (multimodal references) INT8 13.9B 16 GB 19 GB Gaming GPU

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 24 GB of GPU memory.

The license in plain words

A community license that lets you use, modify and distribute MiniMax H3, but only in the "Applicable Territory": the whole world except the European Union, the United Kingdom, South Korea and the United States. Commercial products earning over US$20 million a year need MiniMax's written authorization, every commercial product must show "MiniMax H3" in its interface, and you may not use the model or its outputs to train other AI models. Get legal advice, especially if you or your users are in an excluded region.

  • Attribution or model name must be shown
  • Free commercial use only below a revenue cap
  • Not licensed in some regions
  • Acceptable-use policy applies
  • Outputs must be filtered or reviewed, and AI generation disclosed

More about MiniMax H3 Community

See also

Frequently asked questions

What is HyperFlow H3 Flash?
HyperFlow H3 Flash is a smaller, faster version of the MiniMax H3 video model, made by the Singapore start-up Video Rebirth. It makes a clip of 5 to 15 seconds with matching stereo sound in 4 to 8 steps, and its 9.1 GB GGUF file runs on a 24 GB graphics card. It keeps H3's license, which does not cover the United States, the European Union, the United Kingdom or South Korea.
Can I use HyperFlow H3 Flash commercially?
Yes, with conditions. HyperFlow H3 Flash is released under the MiniMax H3 Community, which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Not licensed in some regions; Acceptable-use policy applies; Outputs must be filtered or reviewed, and AI generation disclosed. Read the license text before you ship a product.
Is HyperFlow H3 Flash open source?
HyperFlow H3 Flash is open-weight, not open source in the strict sense. You can download the model files, but the MiniMax H3 Community adds conditions that the Open Source Initiative definition does not allow.
Can I run HyperFlow H3 Flash on my own computer?
Yes, on a good graphics card. The smallest variant, FL2VA INT8 (text and endpoint frames), needs about 18 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was HyperFlow H3 Flash released?
HyperFlow H3 Flash was released by Video Rebirth on November 4, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
MiniMax H3 MiniMax 33.1B – MiniMax H3 Community (Conditional) 166 GB Big-memory desktop
Pika H3 Flex Pika 33.1B – MiniMax H3 Community (Conditional) 169 GB Big-memory desktop
MiniMax H3-Image MiniMax 33.1B – MiniMax H3-Image Community (Conditional) 155 GB Big-memory desktop
LTX-2.5 LTX (Lightricks) 22B – LTX-2.x Community (Conditional) 16 GB Laptop or phone