Qwen-3.5-INT8-Convrot-Comfy

Qwen3.5 INT8 ConvRot Text Encoders for ComfyUI

INT8 quantized versions of the Qwen3.5 2B, 4B, and 9B vision-language text encoders, converted using ConvRot (Hadamard-rotation outlier suppression) for improved quantization fidelity. All three are fully compatible with ComfyUI's native mixed-precision ops and tested to produce coherent, detailed image descriptions.

Files

Model links

REPO_NAME at main

2B

• qwen3.5_2b_int8_convrot.safetensors

4B

• qwen3.5_4b_int8_convrot.safetensors

9B

• qwen3.5_9b_int8_convrot.safetensors

Model Storage Location

📂 ComfyUI/
├── 📂 models/
│   └── 📂 text_encoders/
│          ├── qwen3.5_2b_int8_convrot.safetensors
│          ├── qwen3.5_4b_int8_convrot.safetensors
│          └── qwen3.5_9b_int8_convrot.safetensors

Usage in ComfyUI

Load with a CLIPLoader node. All three files are drop-in replacements for their BF16 counterparts — no special nodes or settings required. ComfyUI's native Qwen35TEModel_ loader correctly handles INT8 quantized weights through comfy_kitchen's dequant path.

Why ConvRot

ConvRot applies a group-wise Hadamard rotation to weight matrices before INT8 quantization, suppressing the row-wise outliers that are characteristic of DiT and M architectures. This produces sharper, more accurate outputs compared to plain INT8 quantization. These conversions use --convrot-group-size 256, the recommended default per the ConvRot author.

The visual tower in each model is fully protected at native BF16 precision via the --qwen35 preset, which skips first/last layers, embeddings, and all visual components from quantization. Only the core language model transformer layers are quantized, preserving the model's ability to accurately interpret and describe image content.

Quantization Recipe

ctq -i qwen3.5_Xb_bf16.safetensors \
    -o qwen3.5_Xb_int8_convrot.safetensors \
    --int8 --scaling_mode row --simple --low-memory \
    --convrot --convrot-group-size 256 \
    --qwen35 --comfy_quant --save-quant-metadata

Converted using silveroxides/convert_to_quant on an RTX 3070 8GB.

Hardware tested

• RTX 3070 8GB (conversion and inference)
• All three variants load and run successfully within 8GB VRAM with ComfyUI's dynamic VRAM management

Credits

Base models: Qwen/Qwen2.5- Quantization tooling: silveroxides/convert_to_quant ConvRot algorithm: ConvRot paper/implementation