MiniMax H3 - INT4 ConvRot Edition
This is a complete INT4 ConvRot conversion of the original MiniMax H3 model files. The conversion preserves the original model architecture and capabilities while storing its weights at substantially lower precision, making the model easier to run on systems with limited GPU memory.
INT4 ConvRot model
Much smaller model weights
Lower RAM and VRAM usage
Reduced memory-transfer requirements
May load and run faster on compatible hardware
Can introduce small quality differences compared with the original full-precision model
Hugging Face (mirror) - https://huggingface.co/Merserk/MiniMax-H3-INT4-ConvRot
Patreon (support): https://www.patreon.com/mm744
Description
FAQ
Comments (26)
Anyone know if this is safe?
It's safetensors, there is literally nothing that it can execute by itself. At worst, it's not the model that it claims to be and that is fairly simple to ascertain in one run (assuming comfy also supports int4 natively, I don't recall).
This site is more dangerous than a safetensors: https://github.com/civitai/civitai/issues/3434
The file sizes don't match the Comfy-Org models. I would grab it from the HF instead. https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/diffusion_models
These are INT4 ConvRot models, converted for faster performance and lower VRAM usage.
@MM744 Fair enough.
Safe it is , but I would not expect quality. Better to grab from https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/diffusion_models even if it does not fit your VRAM, get the INT8 pruned version. Only get INT4 or NVFP4 if you prepare yourself for disappointment first.
@MM744 I think he meant to ask if it's safe for work, is nsfw! hahah
Can somebody create a minimax h3 workflow for 12gb +16gb ram lowvram specs
it doesn't work for me, it gave me a messy blurry broken video result.
If used, the generated video looks like a Van Gogh work
sorry, tried out t2v, i2v, different resolutions and it looks like s**t like have already said. There might be a separate workflow, but without information: save yourself the download.
Yes painfully slow even with the lowest resolution. 20-25 steps is a lot. should wait for it until it could be usualy at 8 steps or lower like LTXV
there is no way you're gonna get any quality out of these, far too quantized.
Poor quality (too noisy outputs), as always from you
12GB VRAM + 32 RAM = normal at int8 checkpoint)
In fact, H3 needs RAM more than VRAM. The int8 version can even generate 0.5MP, 24fps, 5s in 5 minutes on my setup with 8GB VRAM + 32GB RAM (4060M + DDR5). However, it requires some effort—for example, running it on a Linux environment with no other software open to free up memory. Windows is too bloated. It roughly needs 28GB RAM and 6GB VRAM. My recommended sweet spot parameters are 0.35MP, which can be completed in 3 minutes.
So, go ahead and try int 8! Even at 0.5MP, pushing it up to 10s won't cause OOM, and RAM and VRAM remain basically unchanged, but the time expands to 15 minutes.
Same device reference, wan2.2 at 0.5MP, 20FPS, 4s requires 10 minutes of generation time, and the jelly effect is severe. Open source is so wonderful!
By the way, H3 doesn't need to stack various LoRAs to perform conventional NSFW actions well (I can't think of any unconventional ones). It's both fast and effective.
I am struggling to create even a 5s clear video with my HP Victus15 RTX4060, 16Gig RAM :( . I can create a minute long video with LTX2.3 mxfp8 models. I hope they will do something similar for this newly released model. Tried Q3 but same slow process.
I can confirm. I generate 12 second 0,5 MP @ 24FPS in 16 minutes on my 5060 Ti with 16GB RAM and the Turbo LoRa on the w4a8 model. Wasn't much worse on the INT8 model either, i think it was 10 seconds of 0,5 MP x 24 FPS in some 20 minutes also.
All much better than LTX anyway. And if you factor in the fact that H3 actually adheres to prompts and you can use most generations right off without re-runs while LTX makes you spend all night on the prompt, the saving becomes even more clear.
I wouldn't call running Linux effort though. The opposite :D I am able to watch movies while generating so I am good.
I didn't have that luck.
My setup can generate me 10s vid, 24fps, 640*1056, after 1.5 hours of working with the int8 fl2va model.
But it refuses completetly to generate any video even in the lowest of the lowest quality configs when it comes to the Ref2va models, even with the most aggressively quantized ones.
Always OOm. No cap.
(12Gb Vram+128gb sysRam).
Now I'm not sure if it needs more Ram than Vram
@2600angroup
Before entering the sampling stage, I used the "unload all model" node, which actually unloaded that huge Qwen Clip, thereby saving a huge amount of RAM. Without unloading it in advance, I couldn't even sample at all. I'm not sure whether this also works for VRAM. Hope it's useful to you.
Thanks for the effort, but the output quality is pretty poor, as others already said