---
title: Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI
canonical_url: "https://www.modelscope.cn/models/Abiray/Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI"
md_url: "https://www.modelscope.cn/models/Abiray/Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI.md"
repository: Abiray/Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI
last_updated: 2026-08-04
license: other
base_model:
  - MiniMaxAI/MiniMax-H3
base_model_relation: quantized
parameters: 48.4B
tensor_type:
  - F32
  - F8_E4M3
  - U8
  - I8
  - BF16
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 4690
stars: 32
tags:
  - comfyui
  - nvfp4
  - minimax-h3
  - qwen3-vl
  - text-encoder
  - uncensored
  - video-generation
---

# Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI

> Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI - Abiray 在 ModelScope 开源的模型。Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI This repository contains a direct-from-BF16 NVFP4 (NVIDIA 4-bit Float) quantization of the uncensored Qwen3-VL 32B text encoder for…

Abiray/Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI 是 ModelScope 魔搭社区上的 48.4B 参数机器学习模型，采用 other 许可，基于 MiniMaxAI/MiniMax-H3 构建。

- **Repository**: Abiray/Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI
- **License**: other
- **Parameters**: 48.4B
- **Base model**: MiniMaxAI/MiniMax-H3
- **Tags**: comfyui, nvfp4, minimax-h3, qwen3-vl, text-encoder, uncensored, video-generation
- **Downloads**: 4690
- **Stars**: 32
- **Last updated**: 2026-08-04

Source: https://www.modelscope.cn/models/Abiray/Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI

---

## Qwen3-VL-32B-Heretic-MiniMax-H3-nvfp4-ComfyUI 
   This repository contains a direct-from-BF16 NVFP4 (NVIDIA 4-bit Float) quantization of the uncensored Qwen3-VL 32B text encoder for MiniMax-H3 video generation in ComfyUI. By baking directly from the pristine 65 GB uncompressed BF16 weights (rather than re-quantizing an INT8 build), this version completely avoids double-quantization loss and ConvRot rotation artifacts, yielding a clean, highly responsive ~15.7 GB text encoder that fits on 16 GB and 24 GB GPUs.

# important_download_disclaimer: 
  DO NOT click "Download all files" or clone this entire repository without checking file names!
  This repository contains both the final 4-bit runtime files and the heavy uncompressed intermediate source files.
  
  - qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors (~15.7 GB) : Main NVFP4 Base Encoder (YES - Download this for ComfyUI)
  - base_encoder.safetensors (51.5 GB) : Raw uncompressed BF16 intermediate source (NO - Do not download)
  - tail_encoder.safetensors (15.2 GB) : Raw uncompressed BF16 intermediate source (NO - Do not download)
  
  To save bandwidth: Download ONLY the .safetensors files containing _nvfp4 in their names!

# key_features:
  - "Pristine BF16 Source Bake: Quantized straight from the original 65 GB uncompressed source using convert_to_quant (AdamW AdaRound optimization). No 'double quantization' quality degradation."
  - "16 GB GPU Compatible: The main conditioning encoder fits in ~15.7 GB of VRAM, running dynamically on single 16 GB (RTX 4080 / RTX PRO 2000 / Ada) or 24 GB GPUs (RTX 3090 / 4090 / L4)."
  - "Uncensored (Heretic ARA Edits): Retains the Heretic v1.2.0 ARA edits on language layers 31–40, reducing refusal rates from ~99% to ~4% while preserving full prompt adherence."
  - "Mixed Precision Safeguards: Language blocks use NVFP4 (TensorCoreNVFP4Layout), while embed_tokens and lm_head are kept in INT8 to prevent ComfyUI memory spikes and lookup errors."

# comfyui_setup_and_installation: 
  1. Place the NVFP4 .safetensors file(s) into your ComfyUI directory: ComfyUI/models/text_encoders/MiniMax-H3/
  2. In your workflow, select the file using the CLIPLoader node with type set to 'minimax'.

# hardware_requirements:
  - "GPU: NVIDIA Ada Lovelace, Hopper, or Blackwell architecture (e.g., RTX 40-Series, RTX PRO Ada, L4, L40S) for native NVFP4 hardware acceleration."
  - "Environment: PyTorch 2.8.0+ with CUDA 12.8 / 13.0+ and modern comfy-kitchen nodes."

# acknowledgments_and_credits:
  - "Uncensored Source Weights: llmfan46 (https://huggingface.co/llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic)"
  - "ComfyUI Architecture & Split Concept: ethanfel (https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot)"
  - "Base Architecture: Alibaba Qwen Team (Qwen3-VL-32B-Instruct) and MiniMaxAI (MiniMax-H3)"
  - "Quantization Tooling: silveroxides/convert_to_quant"