MiniMax H3 — W4A8 ConvRot (Pruned)

Quantized ComfyUI-ready checkpoints for MiniMax H3, converted to W4A8 (asym_w4a8_int8 / ConvRot) format with pruning applied, plus a matching quantized text encoder. Built for running MiniMax H3 on consumer GPUs with a much smaller memory footprint than the original BF16/FP16 weights.

Files

File	Description
`minimax_h3_ref2va_pruned-w4a8_convrot_pruned.safetensors`	Reference-to-video/audio (ref2va) diffusion transformer, pruned + W4A8 ConvRot quantized
`minimax_h3_fl2va_pruned-w4a8_convrot_pruned.safetensors`	First-last-to-video/audio (fl2va) diffusion transformer, pruned + W4A8 ConvRot quantized
`minimax_h3_te_w4a8_convrot.safetensors`	Quantized text encoder (Qwen3.5-VL-style), vision tower left unconverted
`Winn_MMH3_FastV2_Unlocked_Pruned-w4a8_convrot_pruned.safetensors`	FastV2 Unlocked, uncensored, pruned diffusion transformer — W4A8 ConvRot quantized
`Winn_MMH3_FastV2_Unlocked_Pruned-int8_convrot_pruned.safetensors`	FastV2 Unlocked, uncensored, pruned diffusion transformer — INT8 ConvRot quantized

Format details

These checkpoints use the asym_w4a8_int8 quantization scheme (ConvRot-rotated INT4 weights + Lloyd-Max codebook + FP8 group scales, running on INT8 GEMM). No calibration data is required at inference time.

Weight bits: 4-bit (ConvRot-rotated, codebook-packed)
Activation/GEMM: INT8
Per-layer metadata: weight_s_rel (FP8 group scale), optional weight_s_channel (FP32 per-channel scale), weight_codebook (Lloyd-Max codebook)
Group size: 16 (default) / ConvRot group size: 256 (default)
Diffusion transformers additionally use a pruned variant to further reduce size

This format requires ComfyUI mainline with asym_w4a8_int8 support in comfy/ops.py / comfy/quant_ops.py (merged into mainline 2026-08-07). No custom branch build is needed.

Text encoder notes

The text encoder is Qwen3.5-VL-style. Its vision tower (visual.* weights) is kept unquantized — only the language-model portion of the text encoder is quantized to W4A8.

Usage

Place the diffusion model file(s) in ComfyUI/models/diffusion_models/.
Place the text encoder in ComfyUI/models/text_encoders/ (or wherever your MiniMax H3 workflow expects it).
Load with the standard MiniMax H3 ComfyUI workflow — the loader will detect the comfy_quant metadata and W4A8 fields automatically.
Requires a recent ComfyUI build (post 2026-08-07) with asym_w4a8_int8 support merged into mainline.

Which diffusion model do I use?

ref2va — use this for reference-image-to-video/audio generation.
fl2va — use this for first-frame/last-frame-to-video/audio generation.
FastV2 Unlocked Pruned variants — pruned, "unlocked" FastV2 diffusion models offered at three quantization levels (W4A8 ConvRot, W4A4, INT8 ConvRot) — pick based on your VRAM budget vs. quality tradeoff, from smallest (W4A4) to highest quality of the three (INT8 ConvRot).

Limitations

Pruned + W4A8 quantization trades some quality for a significantly smaller footprint compared to BF16/INT8 — expect minor degradation versus the full-precision model.
Quantization-aware kernel performance on some INT8 GEMM/dequant paths is still being tuned upstream in comfy-kitchen; speed gains over BF16 may vary depending on your GPU and sequence