berryber09/MiniMax-H3-Fun-Controlnet-Union-2.0-pruned-w4a8

MiniMax-H3 Fun ControlNet-Union 2.0, pruned, W4A8

W4A8 quantization of the MiniMax-H3 Fun ControlNet-Union 2.0 model patch for ComfyUI (the v1 build is at berryber09/MiniMax-H3-Fun-Controlnet-Union-w4a8).

File	minimax_h3_fun_controlnet_union_2.0_pruned_w4a8.safetensors
Size	2.85 GB
Source	alibaba-pai/MiniMax-H3-Fun-Controlnet-Union-2.0, via the pruned bf16 build in Comfy-Org/MiniMax-H3 (minimax_h3_fun_controlnet_union_2.0_pruned_bf16.safetensors)
sha256	7e7d31399f0396819f7aabcf3f7a9cc3e975757570454c5a3bfb093b16859037

Quantization

• Format: asym_w4a8_int8 (4-bit asymmetric weights, int8 activations), ConvRot rotation with group size 256, weight group size 16.
• Applied to the 40 2D projection layers of the 10 control blocks. Norms, embeddings, biases and the AdaLN basis stay in bf16/fp32.
• Quantized from the bf16 build (not from int8). Quantization metadata is stored in the safetensors header (_quantization_metadata, comfy_wxa8), the layout ComfyUI reads natively. The header also keeps the 2.0 loader keys (control_blocks_places, inpaint_masked_pixel_mode=post_norm, minimax_h3_fun_controlnet=adaln_basis).

Usage (ComfyUI)

1. Put the file in models/model_patches/.
2. Load it with ModelPatchLoader (it builds all 10 control blocks), apply with MiniMaxH3FunControlNetApply.
3. Requires a ComfyUI version with the 2.0 loader (10 control blocks, post_norm inpaint mode).

Tested

• Loads and builds 10 control blocks.
• Pose control A/B against the bf16 build on the upstream asset/pose.mp4 example, 6 steps with the DMD ref2va turbo LoRA at 0.6: clean, follows the driving motion.

Notes

• License follows the source model (MiniMax-H3 Community License Agreement).