MiniMax H3 INT4 ConvRot + Mixed FP16/FP32 Video VAE

This repository contains a collection of MiniMax H3 model components intended for lower-memory local inference.

The original MiniMax H3 model and architecture were released by MiniMaxAI:

https://huggingface.co/MiniMaxAI/MiniMax-H3

MiniMax H3 is an omni-modal video generation model capable of generating synchronized video and stereo audio. The official release supports text-to-audio-video, first/last-frame-to-audio-video, and multimodal reference-to-audio-video generation, depending on the selected model variant.

The original MiniMax H3 release, documentation, architecture information, usage instructions and licensing information can be found in the official repository linked above.

INT4 ConvRot model files

The following three quantized model files included in this repository were copied from the MiniMax H3 INT4 ConvRot release by Winnougan:

https://huggingface.co/Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI

minimax_h3_fl2va_pruned-w4a8_convrot_pruned.safetensors
minimax_h3_ref2va_pruned-w4a8_convrot_pruned.safetensors
minimax_h3_te_w4a8_convrot.safetensors

These files use the W4A8 / ConvRot quantization format.

The FL2VA and Ref2VA transformer checkpoints are pruned and W4A8 ConvRot quantized versions of the MiniMax H3 diffusion transformers. The text encoder is also provided in W4A8 ConvRot format, while its visual component remains at higher precision as in the source conversion.

All credit for these INT4 ConvRot conversions belongs to Winnougan and the respective contributors to the ConvRot/W4A8 tooling.

For additional technical information about the quantization format, requirements and original converted files, see the Winnougan repository linked above.

Mixed-precision Video VAE

This repository also contains a modified version of the MiniMax H3 Video / Visual VAE.

The original Video VAE weights originate from the official MiniMax H3 release.

For this repository, I converted the majority of the Video VAE tensors from FP32 to FP16 in order to substantially reduce its memory and storage requirements.

Instead of blindly converting the entire VAE to FP16, selected small and precision-sensitive blocks were intentionally retained in FP32.

The resulting file is therefore a mixed FP16/FP32 Video VAE:

Large weight tensors are stored in FP16.
Selected smaller / precision-sensitive tensors remain FP32.
The goal is to reduce memory usage while preserving as much reconstruction quality and numerical stability as possible.
The model architecture itself was not retrained or fine-tuned.

MiniMax H3 Tubo 4 step loras

Newest Turbo lora step600 was copied from https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora

Model components

FL2VA INT4/W4A8 ConvRot transformer
Ref2VA INT4/W4A8 ConvRot transformer
INT4/W4A8 ConvRot MiniMax H3 text encoder
Mixed FP16/FP32 Video VAE
MiniMax H3 FP32 Audio VAE
Supporting configuration files where required

Intended use

The purpose of this repository is to provide a more practical MiniMax H3 model set for local inference, particularly on consumer GPUs where loading the original full-precision model components can require a very large amount of VRAM and system memory.

The quantized components trade some numerical precision for substantially reduced memory requirements.

The mixed-precision Video VAE is intended to further reduce memory consumption compared with the original FP32 VAE while retaining FP32 precision where it may matter most.

Officially created for use with Comfyui. With some work and the correct (comfyui) loader these also work standalone in other setups.

Credits

Original model

MiniMaxAI — MiniMax H3
https://huggingface.co/MiniMaxAI/MiniMax-H3

INT4 / W4A8 ConvRot conversions

Winnougan — MiniMax-H3-INT4_Convrot_ComfyUI
https://huggingface.co/Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI

Mixed FP16/FP32 Video VAE conversion

Converted from the official MiniMax H3 Video VAE for reduced-memory inference. Most tensors were converted to FP16 while selected small / precision-sensitive tensors were preserved in FP32.

License and usage

This repository contains components derived from or redistributed from other projects.

Users should review and comply with the licensing and usage terms of the original MiniMax H3 model and any applicable terms associated with the converted components.

This repository is not affiliated with or endorsed by MiniMaxAI.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support