AI Potluck
Contribute
Back to Gap Map Model components / Image, video, 3D & music generation

Nunchaku

nunchux-ai
open source / Overall score: 2.0

Nunchaku is an inference engine for diffusion models quantized to 4-bit weights and activations with SVDQuant, a low-rank method that keeps image quality close to the full-precision model while cutting memory several times over. It runs INT4 and NVFP4 builds of FLUX.1, Qwen-Image, SANA, Z-Image and PixArt, and plugs into ComfyUI through a companion node pack. It began in MIT HAN Lab.

The PyPI project named nunchaku is an unrelated data-segmentation library; this engine ships as wheels on its GitHub releases, so no package is declared.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI
source
public(nunchux-ai/nunchaku)
core features withheld
no — no ee, enterprise, commercial or proprietary path

Nunchaku is released under Apache 2.0 with its full source public. Its authors offer consulting and partnership to companies but no separate paid build of the engine.

Adoption

2 low confidence
2.0

The engine is distributed as wheels on its GitHub releases, and the PyPI name it would use belongs to an unrelated project, so GitHub stars are the only comparable signal. A star is not a use.

Capability

2 high confidence
2.0

Nunchaku runs several image model families in 4-bit form, fitting large models such as FLUX.1 onto consumer GPUs. It is a step below stable-diffusion.cpp because it runs image models only, with no video generation.

Verified 2026-09-26