# Nunchaku ComfyUI Plugin

A ComfyUI plugin for Nunchaku, an inference engine that runs 4-bit quantized diffusion models such as Qwen-Image and FLUX with much lower VRAM use.

Nunchaku ComfyUI Plugin brings Nunchaku, an inference engine for 4-bit neural networks quantized with SVDQuant, into ComfyUI. It lets ComfyUI workflows run diffusion models such as FLUX and Qwen-Image at 4-bit precision, cutting VRAM use substantially while aiming to keep image quality close to full precision.

Recent releases added support for 4-bit Qwen-Image-Edit and faster Lightning variants, asynchronous transformer offloading that reduces VRAM use to as little as 3 GiB, native LoRA support with ComfyUI nodes, and INT4 support for older 20-series GPUs, alongside a helper node that installs the correct prebuilt Nunchaku wheel for the user's environment. Pre-quantized models are published on Hugging Face and ModelScope, and a related library provides the underlying quantization tooling.

It is written in Python and released under the Apache License 2.0.

Website: https://nunchaku.tech/docs/ComfyUI-nunchaku/