# Nunchaku

[ICLR2025 Spotlight] SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models

Repository: https://github.com/nunchaku-ai/nunchaku
Canonical: https://ross.abutalabs.com/products/nunchaku
Homepage: https://nunchaku.tech/docs/nunchaku/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: diffusion-models, flux, genai, lora, mlsys, quantization, iclr, iclr2025, comfyui
Last push: 2026-03-07T06:42:49+00:00
Link (homepage): https://nunchaku.tech/docs/nunchaku/
Link (site_page): https://nunchaku.tech/docs/nunchaku/installation/installation.html
Link (site_page): https://nunchaku.tech/docs/nunchaku/installation/setup_windows.html
Link (site_page): https://nunchaku.tech/docs/nunchaku/usage/basic_usage.html
Link (site_page): https://nunchaku.tech/docs/nunchaku/usage/qwen-image.html

## Health v2 (maintenance only)
Score: 65/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 71, release rhythm 67, longevity 47
- inputs: {"age_days": 665, "days_push": 179, "days_rel": 220, "gap_med": 24, "n_releases_24m": 14}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3937, forks 277 (observed 2026-08-28T04:08:30.108187+00:00)

## What it is
Nunchaku is a high-performance inference engine for 4-bit quantized diffusion models (and LLMs) based on the SVDQuant technique from an ICLR 2025 paper. It provides a Diffusers-compatible Python API and a ComfyUI plugin for running models like FLUX.1, Qwen-Image, and SDXL with low VRAM usage.

## Use cases
- run FLUX.1 image generation in 4-bit on consumer GPUs
- generate images with Qwen-Image using less VRAM
- use LoRAs with quantized diffusion models without requantization
- run SDXL text-to-image on low-VRAM GPUs
- use 4-bit diffusion models in ComfyUI workflows
- run image editing models like FLUX.1-Kontext or Qwen-Image-Edit efficiently
- offload diffusion models to CPU to fit in 3-4GB VRAM

## When to choose
- you want to run large diffusion models (FLUX, Qwen-Image, SDXL) on consumer NVIDIA GPUs with limited VRAM
- you need fast 4-bit inference with minimal quality loss
- you use ComfyUI or Diffusers and want drop-in quantized model support
- you want to use off-the-shelf LoRAs with quantized models

## When to avoid
- you need CPU-only or AMD/Intel GPU inference
- you need to quantize your own custom models (use DeepCompressor instead)
- you need maximum quality at full precision
- you work with non-supported model architectures

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, image-processing, gpu-computing
- domain: deep-learning, image-processing, machine-learning, gpu-computing
- platform: python, windows
- tags: diffusion-models, quantization, svdquant, 4-bit, flux, lora, comfyui, inference-engine, text-to-image, linux, gpu

## Member repositories
- nunchaku-ai/nunchaku (main) score 65
- nunchaku-ai/ComfyUI-nunchaku (plugin) score 62

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:08:30.108187+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T18:24:38.624433+00:00, confidence not recorded.
  - readme: https://github.com/nunchaku-ai/nunchaku (fetched 2026-08-28T04:08:30.108187+00:00, sha 2eb77ea64b29)
  - homepage: https://nunchaku.tech/docs/nunchaku/ (fetched 2026-08-29T09:18:35.945726+00:00, sha f8f2b6399bc8)
  - site_page: https://nunchaku.tech/docs/nunchaku/installation/installation.html (fetched 2026-08-29T09:18:35.954573+00:00, sha 03ff90dbad67)
  - site_page: https://nunchaku.tech/docs/nunchaku/installation/setup_windows.html (fetched 2026-08-29T09:18:35.956690+00:00, sha d417a05fbfe3)
  - site_page: https://nunchaku.tech/docs/nunchaku/usage/basic_usage.html (fetched 2026-08-29T09:18:35.958482+00:00, sha c781ffcee3d0)
  - site_page: https://nunchaku.tech/docs/nunchaku/usage/qwen-image.html (fetched 2026-08-29T09:18:35.960181+00:00, sha e9ccf136364d)
  - site_page: https://nunchaku.tech/docs/nunchaku/usage/qwen-image-edit.html (fetched 2026-08-29T09:18:35.962257+00:00, sha 3e37684c0126)
  - site_page: https://nunchaku.tech/docs/nunchaku/usage/lora.html (fetched 2026-08-29T09:18:35.964246+00:00, sha c70e48102678)
  - site_page: https://nunchaku.tech/docs/nunchaku/usage/kontext.html (fetched 2026-08-29T09:18:35.966625+00:00, sha 29f3b7874770)
  - site_page: https://nunchaku.tech/docs/nunchaku/usage/sdxl.html (fetched 2026-08-29T09:18:35.968763+00:00, sha 2d95c44d1db5)
- Data as of 2026-08-30T08:39:29.467469+00:00.