Qwen3 VL 32B Heretic MiniMax H3 NVFP4

Same size as Comfy-Org's official NVFP4 encoder, so it is a drop-in replacement: point CLIPLoader at this file instead and the rest of your MiniMax-H3 workflow is unchanged.

Why

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.

Measured

Generated 6 s of 480×864 vertical video with audio, MiniMax-H3 fl2va pruned INT8 diffusion model, res_multistep 20 steps, ComfyUI 0.30.0, Sage Attention on:

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.

Files

qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors   15.7 GB

Place in ComfyUI/models/text_encoders/ and select it in CLIPLoader (type: minimax). You still need the diffusion model and VAEs from Comfy-Org/MiniMax-H3.

How it was made (and the trap you must avoid)

The upstream weights are rotated. Its comfy_quant metadata 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 NVFP4 without undoing the rotation, you get a file that loads, runs, and produces completely 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 by H again 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) is left 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-layer comfy_quant metadata, so the mixed file loads with no special handling.

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

Provenance

• 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.

Notes

• 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.