license: other license_name: personal-entertainment-use-only base_model:

• MiniMaxAI/MiniMax-H3
• Comfy-Org/MiniMax-H3 tags:
• comfyui
• minimax-h3
• qwen3-vl
• text-encoder
• safetensors
• int8
• convrot
• quantized
• uncensored

qwen3vl_32b_minimax_h3_int8_convrot_uncensored-by-linjian257

中文 | English

模型简介

这是一个供 MiniMax-H3 ComfyUI 工作流使用的 Qwen3-VL-32B INT8 TensorWise + ConvRot 社区修改版文本编码器。该版本由 linjian257 从修改后的 BF16 权重量化得到，重点是在保持 ComfyUI 工作流兼容性的同时明显降低权重文件体积。

本文件只是 MiniMax-H3 的文本编码器组件，并不是完整的视频生成模型，也不是可直接在 LM Studio、llama.cpp 或 vLLM 中作为聊天模型加载的完整 Qwen3-VL 模型。

关键信息

量化信息

• 算法：int8_tensorwise，版本 0.1.0
• 存储类型：signed INT8 权重
• 缩放方式：每输出通道一个 FP32 scale（per_channel / out_features）
• Scale 计算：绝对值最大值（amax）
• 舍入：nearest-even
• 量化范围：model.* 与 visual.* 中选定的二维权重；bias 等不适合量化的张量保持原精度
• ConvRot：启用，group size 为 256

该文件是从 linjian257 制作的 BF16 社区修改版权重重新量化得到的衍生版本，并非官方 qwen3vl_32b_minimax_h3_int8_convrot.safetensors 的改名或逐字节镜像。

“Uncensored”用于区分版本，不代表对所有提示词、语言、输入形式或工作流效果作出保证。

安装与使用

1. 使用支持 MiniMax-H3 及原生 INT8 ConvRot 的较新版本 ComfyUI。如果出现未知量化格式、缺少 weight_scale 或 ConvRot 相关错误，请先更新 ComfyUI 及其依赖。

2. 将文件放入：

   ComfyUI/models/text_encoders/
   

3. 下载 MiniMax-H3 工作流所需的扩散模型和 VAE 文件。官方整理版见 Comfy-Org/MiniMax-H3。

4. 打开官方 T2V、I2V 或 R2V 工作流。

5. 在 CLIPLoader 节点中选择本文件，并保持模型类型为 minimax。

如果 ComfyUI 找不到该文件，请确认它位于 models/text_encoders 而不是 models/checkpoints 或 models/diffusion_models，然后刷新模型列表或重启 ComfyUI。

已知限制

• 本仓库只提供文本编码器权重；运行 MiniMax-H3 还需要匹配的扩散模型、视频 VAE，以及在需要音频时使用的音频 VAE。
• INT8 通常能减少权重存储和内存压力，但实际显存占用、速度与兼容性取决于 GPU、后端、Torch/ComfyUI 版本、卸载设置和完整工作流；本模型卡不承诺特定加速倍数。
• ConvRot 需要运行时正确识别并应用对应旋转元数据。较旧或不兼容的加载器可能报错，也可能错误加载并产生异常输出。
• INT8 量化和社区修改可能相对原始 BF16 版本带来细微质量变化，包括提示词遵循、稳定性、事实性或生成质量方面的差异。
• 当前未随本模型卡提供系统化的 BF16/官方 INT8 对比、安全评测或端到端视频质量基准。
• 该权重不是官方 Comfy-Org 文件的原样镜像；请勿使用官方文件哈希校验本衍生版本。

安全与责任

该版本可能生成冒犯性、危险、违法、误导或不适宜的内容。使用者必须自行评估输出、遵守适用法律和平台规则，并为生成、发布和传播的内容负责。请勿将其用于伤害未成年人、侵犯隐私、欺诈、仇恨、军事或其他违法有害用途。

使用许可

本权重采用个人娱乐使用许可 v1.0：仅允许个人娱乐、个人学习、非商业研究和非商业测试。未经 linjian257 书面许可，不得用于商业用途、收费服务、商业部署、转售、再分发或转授权。

致谢

• MiniMaxAI — MiniMax-H3
• Qwen — Qwen3-VL-32B
• Comfy-Org — ComfyUI 适配、INT8 运行支持及权重整理
• linjian257 — 本衍生版本的处理、量化与发布

English

Model description

This is a community-modified and quantized INT8 TensorWise + ConvRot Qwen3-VL-32B text encoder intended for MiniMax-H3 workflows in ComfyUI. It was produced by linjian257 from modified BF16 weights, with the goal of substantially reducing the weight-file size while retaining compatibility with the ComfyUI workflow.

This file is only the text-encoder component of MiniMax-H3. It is not a complete video-generation model and is not a standalone Qwen3-VL chat model that can be loaded directly in LM Studio, llama.cpp, or vLLM.

Key facts

Quantization details

• Algorithm: int8_tensorwise, version 0.1.0
• Storage: signed INT8 weights
• Scaling: one FP32 scale per output channel (per_channel / out_features)
• Scale method: absolute maximum (amax)
• Rounding: nearest-even
• Selection: eligible 2D weights under model.* and visual.*; unsuitable tensors such as biases remain at their source precision
• ConvRot: enabled with group size 256

This is a newly quantized community derivative of the modified BF16 weights produced by linjian257. It is not a rename or byte-identical mirror of the official qwen3vl_32b_minimax_h3_int8_convrot.safetensors file.

“Uncensored” distinguishes this release and does not guarantee identical behavior across all prompts, languages, input types, or workflows.

Installation and usage

1. Use a recent ComfyUI build with MiniMax-H3 and native INT8 ConvRot support. If you see unknown quantization-format, missing weight_scale, or ConvRot-related errors, update ComfyUI and its dependencies first.

2. Place the file in:

   ComfyUI/models/text_encoders/
   

3. Download the matching MiniMax-H3 diffusion model and VAE files. The official ComfyUI repack is available at Comfy-Org/MiniMax-H3.

4. Open the official T2V, I2V, or R2V workflow.

5. Select this file in the CLIPLoader node and keep the model type set to minimax.

If the file is not listed, verify that it is under models/text_encoders, not models/checkpoints or models/diffusion_models, then refresh the model list or restart ComfyUI.

Limitations

• This repository provides only the text encoder. MiniMax-H3 also requires a compatible diffusion model, video VAE, and an audio VAE when audio generation is used.
• INT8 generally reduces weight storage and memory pressure, but actual VRAM use, speed, and compatibility depend on the GPU, backend, Torch/ComfyUI version, offloading settings, and complete workflow. No specific speedup is guaranteed.
• ConvRot requires a loader that recognizes and applies the rotation metadata correctly. Older or incompatible loaders may fail or may produce invalid output.
• INT8 quantization and community modifications can introduce small quality changes relative to the original BF16 version, including differences in prompt adherence, stability, factuality, or generation quality.
• No systematic BF16/official-INT8 comparison, safety evaluation, or end-to-end video-quality benchmark is provided with this model card.
• This is a modified derivative, not a byte-identical mirror of the official Comfy-Org file. Do not validate it against the official file hash.

Safety and responsibility

This derivative may produce offensive, dangerous, illegal, misleading, or otherwise inappropriate content. Users are responsible for reviewing outputs, complying with applicable law and platform policies, and controlling how generated material is used or distributed. Do not use it to harm minors, violate privacy, commit fraud, promote hate, support military purposes, or engage in other illegal or harmful activity.

License

These weights are released under the Personal Entertainment Use License v1.0. They may be used only for personal entertainment, personal learning, non-commercial research, and non-commercial testing. Commercial use, paid services, commercial deployment, resale, redistribution, and sublicensing require prior written permission from linjian257.

Credits

• MiniMaxAI — MiniMax-H3
• Qwen — Qwen3-VL-32B
• Comfy-Org — ComfyUI integration, INT8 runtime support, and repackaged weights
• linjian257 — modification, quantization, and release of this derivative