HyperFlow H3 Flash in ComfyUI: 13.7B MiniMax H3 GGUF Setup, Files and Settings

HyperFlow H3 Flash is a slimmed-down MiniMax H3 video model from Video Rebirth. They cut the transformer from about 31B parameters to 13.7B, and it renders a clip in 4, 6 or 8 steps at CFG 1, with 32 kHz stereo audio included.

The GGUF files start at 7.2 GB, so it is the first H3 build many people with a single 24 GB card can realistically try. It was released on 6 November 2026, and it is not an official MiniMax release.

This guide covers every file you need with its exact ComfyUI folder, the install, the four workflow templates, the settings that matter, honest speed numbers, the licence catch, and fixes for the errors people are hitting.

What HyperFlow H3 Flash Is

MiniMax H3 is one of the strongest open video models right now, but its transformer is roughly 31B parameters. That is heavy even on a 32 GB card. If you are new to it, my MiniMax H3 ComfyUI guide covers the full model and every workflow.

Video Rebirth shrank that transformer with low-rank factorisation: big weight matrices are replaced by pairs of thinner ones that approximate them. The result is a 13.7B transformer for text and frame tasks (they call it fl2va) and a 13.9B one for reference-to-video (ref2va).

Everything else is stock H3. You keep the same Qwen3-VL text encoder, the same video VAE, the same audio VAE and the same native H3 nodes in ComfyUI. Only the diffusion model and its loader and scheduler change.

It still makes sound. Every clip comes out with 32 kHz stereo audio, 5.2 to 15 seconds long, at 24 fps and 1344×768 by default.

What it can do

• Text-to-video from a prompt.
• First frame (image-to-video).
• First and last frame (the model fills the motion in between).
• Reference-to-video with up to 9 images in ComfyUI. The Python path takes up to 12 images, plus video and audio references.

Files, Sizes and Which GGUF to Pick

You need five files. The two diffusion GGUFs come from VideoRebirthAI/hyperflow-h3-flash. The encoder and VAEs come from Comfy-Org/MiniMax-H3, and if you already run H3 you have them.

The GGUFs sit in the fl2va_gguf/ and ref2va_gguf/ folders of the Hugging Face repo. Here is how the quant levels compare:

If you are not sure what Q4_K_M or Q8_0 mean, my GGUF quantization guide explains it in plain words. For H3 Flash, I would start with Q4_K_M and only move to Q8_0 if you have VRAM left over.

Install in ComfyUI (Step by Step)

1. Update ComfyUI to 0.43.0 or newer

Native H3 support, including the MiniMaxH3ImageToVideo and MiniMaxH3ReferenceToVideo nodes, needs ComfyUI 0.43.0. On the portable build, run update\update_comfyui.bat. On a git install, run git pull in the ComfyUI folder.

2. Install the two node packs

Open a terminal in ComfyUI/custom_nodes and run:

git clone https://github.com/city96/ComfyUI-GGUF
git clone https://github.com/Video-Rebirth/ComfyUI-HyperFlow-Flash

Then install the HyperFlow requirements with ComfyUI’s own Python. On a normal install:

pip install -r ComfyUI-HyperFlow-Flash/requirements.txt

On the Windows portable build, use the embedded Python from the ComfyUI_windows_portable folder:

python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-HyperFlow-Flash\requirements.txt

ComfyUI-GGUF also needs the gguf Python package. If its nodes show red, install its requirements the same way. My ComfyUI-GGUF guide has the details.

3. Download the files into the right folders

Use the table above. The two VAEs both go in models/vae/, and the text encoder goes in models/text_encoders/.

4. Restart and open a template

Restart ComfyUI, then go to Workflow > Browse Templates > ComfyUI-HyperFlow-Flash. Pick a template, check that each loader points at your files, and press Run.

The Four Workflow Templates

Each template swaps two things compared with a normal H3 workflow. The HyperFlow GGUF Loader replaces the H3 model loader, and HyperFlow Sigmas replaces the scheduler. The rest of the graph is the standard H3 layout, so you can copy prompts and conditioning from your existing workflows.

The loader has a weights_in setting with three choices: auto, disk and ram. Leave it on auto unless you are fighting memory (see Troubleshooting).

Best Settings

Because there is no negative prompt, put everything into the positive prompt: subject, action, camera move and sound. H3 responds well to describing the audio too, for example “soft rain on a window, distant thunder”.

Real Speed: What to Expect

These are Video Rebirth’s own numbers for a 5.2-second clip at 6 steps:

They also claim 12.7x faster than the H3 baseline on a single B200. Be careful with that number: the baseline is the original many-step H3, not the fast H3 setups most ComfyUI users already run.

The community results are more modest. Kijai measured roughly 10% faster than the pruned full model inside ComfyUI, and LukeG89 said it was not worth switching. I have not benchmarked it on my own rig yet, so treat all of these as reported figures.

The way I read it, the main win is memory, not speed. Half the parameters means it fits more easily, and that matters more than a few seconds per clip if you are on 24 GB.

VRAM and System RAM

• 24 GB VRAM is the smallest setup that has been tested (Q4_K_M).
• 16 GB cards are not measured yet. Q3_K_M might fit with offloading, but nobody has published results.
• System RAM runs out first. The 15.7 GB text encoder plus the GGUF plus the VAEs add up quickly. On a PC with 16 or 32 GB of RAM, start ComfyUI with --fast-disk:

python main.py --fast-disk

On the Windows portable build, add --fast-disk to the end of the line in run_nvidia_gpu.bat. If you are choosing a card for this kind of work, my RTX 5090 vs RTX 5080 for AI comparison may help.

Flash vs Full H3 vs the 8-Step LoRA

Video Rebirth also makes the HyperFlow 8-step LoRA, which runs on the full H3. If you rely on character LoRAs, camera LoRAs or any of the popular H3 LoRAs, stay on full H3 with that LoRA. If you just need H3 to fit at all, use Flash.

Licence: Read This First

The node code is GPL-3.0. The weights are a different story: they count as a Model Derivative of MiniMax-H3, so they fall under the MiniMax H3 Community License Agreement.

The README says that agreement does not cover use in the European Union, the United Kingdom, the Republic of Korea or the United States without MiniMax’s authorisation, and it comes with an Acceptable Use Policy.

If you live in one of those places, or plan to use clips commercially, read the licence on the Hugging Face page before you download anything.

Troubleshooting

Red nodes: MiniMaxH3ImageToVideo or MiniMaxH3ReferenceToVideo missing

Your ComfyUI is older than 0.43.0. Update it, then restart.

HyperFlow GGUF Loader is red or fails to import

Install ComfyUI-GGUF too, and run the HyperFlow requirements.txt with ComfyUI’s own Python, not your system Python. Check the console at startup for the actual import error.

“Unknown model architecture!” when loading the GGUF

You loaded the file with ComfyUI-GGUF’s own Unet Loader (GGUF). That loader only accepts architectures on its built-in list, and MiniMax H3 is not on it. HyperFlow’s GGUFs also carry no architecture tag on purpose.

Use the HyperFlow GGUF Loader node instead (it is already in the templates). It reads the tensors itself and only borrows ComfyUI-GGUF’s dequantising code, which is why the stock city96 ComfyUI-GGUF is the right one to install. You do not need a fork.

My H3 LoRA does nothing or throws a shape error

That is expected. Turbo LoRAs, sliders and character LoRAs trained on the full 31B H3 do not match the factorised layers. Remove them.

ComfyUI crashes or freezes while loading

That is usually system RAM, not VRAM. Start ComfyUI with --fast-disk, close other apps, and try setting weights_in to disk on the loader.

CUDA out of memory during sampling

Drop to Q3_K_M, shorten the clip towards 5.2 s, or keep the default 1344×768 instead of going bigger.

Washed-out or burnt video

Check that CFG is exactly 1.0 and the sampler is euler. Raising CFG on a model trained at CFG 1 is the most common cause.

Templates do not appear in the browser

The node pack did not load. Look for an import error in the console, then reinstall the requirements.

My Take

HyperFlow H3 Flash is not the huge speed jump the “12.7x” headline suggests, at least not in ComfyUI on consumer cards. What it does well is fit H3 into far less memory while keeping native audio and all four tasks.

If you have a 24 GB card and the full model keeps running you out of RAM, it is worth trying today. If you depend on LoRAs, or you are in the EU, UK, Korea or the US and need clear licence terms, I would wait or stick with the full model.

FAQ

Is HyperFlow H3 Flash made by MiniMax?

No. It comes from Video Rebirth and is a derivative of MiniMax H3, not an official MiniMax release.

Can I run HyperFlow H3 Flash on 16 GB VRAM?

Nobody has measured it yet. 24 GB is the smallest tested setup. Q3_K_M (7.2 GB) with offloading may work, but expect it to be slow and to need plenty of system RAM.

Does it generate audio?

Yes. Every clip includes 32 kHz stereo sound, the same as full H3.

Which GGUF should I download?

Q4_K_M for a 24 GB card. Pick Q8_0 if you have VRAM to spare and want the closest to full quality, or Q3_K_M if you are short on memory.

Do my MiniMax H3 LoRAs work with it?

No. LoRAs trained on the full 31B model will not load on the factorised 13.7B transformer.

What steps and CFG should I use?

Euler, 6 steps and CFG 1.0. Use 4 steps for quick drafts and 8 for final renders. Leave the negative prompt empty.

Where do the files go?

The GGUFs go in ComfyUI/models/diffusion_models/, the Qwen3-VL encoder in ComfyUI/models/text_encoders/, and both VAEs in ComfyUI/models/vae/.

Can I use the videos commercially?

Check the MiniMax H3 Community License first. It has an Acceptable Use Policy and excludes some regions without MiniMax’s authorisation.

Sources

• Hugging Face: VideoRebirthAI/hyperflow-h3-flash
• GitHub: Video-Rebirth/ComfyUI-HyperFlow-Flash (README, file table, licence)
• GitHub: Video-Rebirth/hyperflow-h3-flash (code and benchmarks)
• Hugging Face: Comfy-Org/MiniMax-H3 (text encoder and VAEs)
• ComfyUI Wiki: HyperFlow H3 Flash news