sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4 · Hugging Face

Source: https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4

https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4#qwen3-vl-32b-heretic-minimax-h3-text-encoder–nvfp4Qwen3-VL-32B Heretic (MiniMax-H3 text encoder) — NVFP4

The uncensored MiniMax-H3 text encoder, in 15.7 GB — it fits on a single 16 GB card.

This is a mixed-precisionNVFP4re-quantization ofethanfel/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot, the Heretic (uncensored) text encoder forMiniMax-H3video generation.

buildsizefits 16 GB card?Heretic INT8-ConvRot (upstream)26.4 GBno (needs offload)Heretic NVFP4 (this repo)15.7 GByesComfy-Org NVFP4 (censored)15.7 GByes Same size as Comfy-Org’s official NVFP4 encoder, so it is adrop-in replacement: pointCLIPLoaderat this file instead and the rest of your MiniMax-H3 workflow is unchanged.

https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4#whyWhy

80 GB cards always had the choice. This is for the people who don’t have one. Creative work shouldn’t require datacenter hardware — that’s the whole point of quantizing it.

https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4#measuredMeasured

Generated 6 s of 480×864 vertical video with audio, MiniMax-H3fl2vapruned INT8 diffusion model,res\_multistep20 steps, ComfyUI 0.30.0, Sage Attention on:

valueGPU1× RTX PRO 2000 Blackwell (16 GB, sm_120)peak VRAM during generation**~9.9 GB**encoder staged in VRAM14.9 GB (dynamic loading)system RAM used by the ComfyUI process~36 GB Output was compared against the upstream INT8-ConvRot build on identical prompt and seed. The two are visually equivalent — the quantization does not change what the encoder will describe.

https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4#filesFiles

qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors   15.7 GB

Place inComfyUI/models/text\_encoders/and select it inCLIPLoader(type:minimax). You still need the diffusion model and VAEs fromComfy-Org/MiniMax-H3.

https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4#how-it-was-made-and-the-trap-you-must-avoidHow it was made (and the trap you must avoid)

**The upstream weights are rotated.**Itscomfy\_quantmetadata reads:

{"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256, "per_row": true}

ConvRot stores each weight already multiplied by a normalized Hadamard matrix (W\_stored = W @ Hᵀ, per 256-wide group) and rotates activations to match at runtime. If you dequantize those weights and re-quantize them to plain NVFP4without undoing the rotation, you get a file that loads, runs, and producescompletely unrelated video— in our first attempt a prompt for a war-torn street rendered a speedboat on open water. Nothing errors. The conditioning is simply meaningless.

Because the Hadamard is orthogonal, the fix is to multiply byHagain before re-quantizing:

from comfy_kitchen.backends.eager.convrot_w4a4 import _build_hadamard

def unrotate(w, gs=256):                      # w: dequantized [out, in]
    out_f, in_f = w.shape
    h = _build_hadamard(gs, device=w.device, dtype=torch.float32).to(w.dtype)
    return torch.matmul(w.reshape(out_f, in_f // gs, gs), h).reshape(out_f, in_f)

Mixed precision.350 linear layers are NVFP4 (TensorCoreNVFP4Layout, group size 16).model\.embed\_tokens(151936 × 5120 = 778 M params) isleft as INT8— the same choice Comfy-Org made in the official NVFP4 build. Quantizing it gives little size benefit and it is the one layer whose temporaries will OOM a 16 GB card during baking. ComfyUI reads per-layercomfy\_quantmetadata, so the mixed file loads with no special handling.

The bake runs on a single 16 GB GPU in about two minutes.

https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4#provenanceProvenance

• Uncensoring / Heretic work:ethanfel
• Original encoder:Qwen3-VL-32B(Alibaba / Qwen team), as adapted for MiniMax-H3
• MiniMax-H3:MiniMaxAI
• ComfyUI packaging conventions and quantization layouts:Comfy-Org
• This NVFP4 re-quantization:Lna-Lab(@Tono_Ken3)

Licensing follows the upstream repositories; the MiniMax-H3 model weights themselves are subject to the MiniMax H3 Community License.

https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4#notesNotes

• Re-quantizing INT8 → NVFP4 means this build inherits the upstream INT8 rounding. A bake straight from BF16 would be marginally cleaner; we did not have those weights.
• Blackwell (sm_120) was used for both baking and inference. NVFP4 needs hardware support.
• If output looks unrelated to your prompt rather than merely lower quality, suspect a rotation mismatch, not the quantization.