martin-rizzo/Qwen3-4B-INT8-ConvRot-ComfyUI

Qwen3-4B INT8-ConvRot (Z-Image / ComfyUI)

This repository provides an INT8-ConvRot quantized version of the Qwen3-4B model. It has been specifically tailored for use as a Text Encoder for the Z-Image generation model in ComfyUI or any other image generation application.

The ConvRot (Rotation of Convolutions/Activations) technique mitigates outliers prior to INT8 quantization. This process enables a significant reduction in VRAM usage while maintaining high fidelity in prompt comprehension compared to the original FP16 weights.

Model Specifications

• Parameters: 4.0B (3.6B Non-Embedding)
• Layers: 36
• Native Context Length: 32,768 tokens

Provided Files

• qwen3-4b_int8_convrot_fp16emixed.safetensors: Quantized model weights in INT8 format, utilizing mixed FP16 precision in specific layers.

Installation Instructions for ComfyUI

To utilize this checkpoint as a Text Encoder within your workflow, follow these steps:

1. Download the qwen3-4b_int8_convrot_fp16emixed.safetensors file.
2. Place the file into the corresponding text encoder directory of your ComfyUI installation. The path is typically: ComfyUI/models/text_encoders/
3. In the ComfyUI interface, select this file within the appropriate Text Encoder loader node.

Production Details

This quantized file was generated from the original checkpoint provided by Qwen (https://huggingface.co/Qwen/Qwen3-4B). The conversion was performed using the z2comfy utility with the following command:

./z2comfy.sh Qwen3-4B --int8convrot --scales-dtype FP32 --scales-trials 32 --mixed-dtype FP16E --mixed-small

Note: The z2comfy command is an experimental utility developed for personal use. It is not fully polished yet and may present some issues, but it is available for download at: https://github.com/martin-rizzo/ZImageCkptTools

References

• Base Model: Qwen/Qwen3-4B
• Tools: ZImageCkptTools