starsfriday/MiniMax-H3-w4a8

MiniMax H3 W4A8 for ComfyUI

Mixed W4A8 checkpoints for MiniMax H3 Ref2VA and FL2VA, converted from the pruned BF16 ComfyUI checkpoints.

These checkpoints require the custom loader and comfy-kitchen W4A8 build from starsFriday/ComfyUI-W4A8-Loader. Do not use ComfyUI's standard Load Diffusion Model node; it does not register the experimental asym_w4a8_int8 format.

Files

File	Variant	Size
minimax_h3_ref2va_pruned_w4a8_mixed.safetensors	Ref2VA reference image, video, and audio workflows	12.54 GB
minimax_h3_fl2va_pruned_w4a8_mixed.safetensors	FL2VA text/image and first/last-frame workflows	12.54 GB

Installation

Stop ComfyUI, activate the Python environment used to run it, then install the node pack and native W4A8 implementation:

cd ComfyUI/custom_nodes
git clone https://github.com/starsFriday/ComfyUI-W4A8-Loader.git
cd ComfyUI-W4A8-Loader
python install_comfy_kitchen_w4a8.py

Download either checkpoint into one of these directories:

ComfyUI/models/unet/
ComfyUI/models/diffusion_models/

Restart ComfyUI and use:

advanced/loaders → Load Experimental W4A8 Diffusion Model

In an existing MiniMax H3 workflow, replace only the standard diffusion-model loader with this node. Keep the remaining VAE, text encoder, conditioning, sampler, and output connections unchanged.

Requirements

• Linux and an NVIDIA GPU with compute capability SM 80 or newer
• Python 3.10 or newer
• A CUDA-enabled PyTorch build using CUDA 12.8 or newer
• CUDA Toolkit 12.8 or newer, including nvcc, to build comfy-kitchen
• A recent ComfyUI version; tested with ComfyUI 0.30.0, Python 3.12.9, PyTorch 2.9.1+cu130, CUDA Toolkit 13.1, and RTX 4090

Windows, AMD/ROCm, Apple Silicon, and CPU-only execution are not supported by the provided installer.

Quantization

For each of the 50 MiniMax H3 transformer blocks, the converter quantizes these four weights:

• attn.qkv_proj.weight
• attn.out_proj.weight
• mlp.fc1.weight
• mlp.fc2.weight

The 200 targeted weights use packed 4-bit ConvRot storage, group size 16, ConvRot group size 256, a calibration-free 16-value codebook, and FP8 E4M3 scales. The optimized CUDA path quantizes activations to INT8 at runtime and uses INT8 GEMM. Non-targeted tensors remain in their source dtype, so these are mixed-precision checkpoints.

Each pruned BF16 source checkpoint is 40.23 GB; its mixed W4A8 result is 12.54 GB, a 68.8% file-size reduction. W4A8 is lossy, and file-size reduction does not guarantee an equal reduction in peak VRAM or workflow runtime.

Source and license

• Base model: MiniMaxAI/MiniMax-H3
• ComfyUI BF16 checkpoints: Comfy-Org/MiniMax-H3
• Loader, converter, installer, and usage documentation: starsFriday/ComfyUI-W4A8-Loader
• W4A8 native implementation: Comfy-Org/comfy-kitchen

This is an independent experimental conversion, not an official MiniMaxAI or ComfyUI release. Use of the weights remains subject to the MiniMax H3 Community License Agreement.