MiniMax H3 w4a8: Kijai's Experimental 4-Bit Weights Land in ComfyUI
Kijai publishes experimental w4a8 quantized MiniMax H3 weights (12.5 GB FL2VA) with an int8 convrot VAE. ComfyUI core support merged in PR #15308.
Kijai published experimental w4a8 quantized MiniMax H3 weights on Hugging Face, paired with an int8 convrot VAE. The asymmetric 4-bit layout stores weights at about 0.56 bytes per element and runs them through int8 GEMM, cutting the pruned FL2VA checkpoint down to 12.5 GB (repository).
The format is backed by ComfyUI core work: Comfy-Org/ComfyUI #15308 "Support asym w4a8_int" merged on August 7, and the int8 convrot VAE support in Comfy-Org/ComfyUI #15334 merged on August 6.
This is a work-in-progress release for testing only. Treat the quality and speed numbers as early data points, not final results.
Files
| File | Size | Role |
|---|---|---|
minimax_h3_fl2va_pruned_w4a8_mixed.safetensors | 12.5 GB | FL2VA pruned diffusion model (text-to-video, first/last-frame) |
minimax_h3_ref2va_pruned_w4a8_mixed.safetensors | 11.8 GB | Ref2VA pruned diffusion model (reference-conditioned) |
minimax_h3_video_vae_int8_convrot.safetensors | 3.2 GB | int8 convrot video VAE, roughly 1.5x faster than the fp16 VAE |
How the w4a8 layout works
The layout comes from comfy-kitchen PR #90 ("Add optimized w4a8 with int8 codebook"), a calibration-free 4-bit weight storage scheme that runs on int8 GEMM with ConvRot activations:
- ConvRot-rotated int4 weights plus a per-tensor Lloyd-Max codebook and fp8 group scales
- Dequantized to grouped int8 and fed into an int8 CUTLASS GEMM
- ~0.073 weight relL2 error on real DiT weights (NVFP4 measures ~0.094)
- ~0.56 bytes per element, ~1.09x faster than int8, no calibration pass needed
- Backends: CUDA (chunked fused int4-to-int8 dequant + strided INT8 GEMM) plus Triton and pure-torch eager paths for AMD/CPU
A frame from Kijai's example w4a8 result on an RTX 5090 (shared in comfy-kitchen PR #90)
Using it in ComfyUI
Run the w4a8 diffusion model with the official H3 workflow templates (video_minimax_h3_t2v and friends), swapping the model loader to the w4a8 checkpoint and the VAE loader to minimax_h3_video_vae_int8_convrot.safetensors. The int8 convrot VAE requires the merged ComfyUI core support from PR #15334.
Availability
- Weights: Kijai/MiniMax-H3-experimental on Hugging Face
- Core support: ComfyUI #15308 (w4a8) and ComfyUI #15334 (int8 convrot VAE)
- Quantization backend: comfy-kitchen #90
Comments
Sign in with GitHub to join the discussion.