Hugging Face has integrated Nunchaku into Diffusers, enabling 4-bit diffusion inference through the library. Nunchaku is an inference engine for quantized diffusion models, and the new integration allows Diffusers users to load a Nunchaku transformer in place of the standard model component.
The blog demonstrates this with a FLUX.1-dev pipeline, where the transformer is replaced by a Nunchaku quantized version. The main benefits are a smaller memory footprint and faster inference, while keeping the familiar Diffusers API. The post is a single-source announcement, so it does not include independent comparisons with other quantization approaches.