# Hugging Face — Bringing Nunchaku 4-bit Diffusion Inference to Diffusers

- Company: Hugging Face (huggingface.co)
- Announced: 2026-07-23T00:00:00+00:00
- Category: developer-tool-release
- Subject: Platform
- Models affected: ERNIE-Image-Turbo, FLUX, Qwen-Image
- Source: https://huggingface.co/blog/nunchaku-diffusers
- Record: https://forck.live/items/1452-bringing-nunchaku-4-bit-diffusion-inference-to-diffusers

Hugging Face integrates Nunchaku 4-bit diffusion inference into Diffusers, introducing Nunchaku Lite for native loading of quantized checkpoints without a separate inference engine, enabling faster inference and lower memory usage on compatible GPUs, including Blackwell-based NVIDIA GPUs for NVFP4 variants. The integration uses the Diffusers library and the kernels package, making it simple to load pre-quantized models with from_pretrained(). Benchmarks show about 30% speedup and significant VRAM reduction compared to BF16 pipelines. The diffuse-compressor toolkit allows users to quantize new architectures and publish them as standard Diffusers repositories.

## Evidence

Verbatim from https://huggingface.co/blog/nunchaku-diffusers:

> With current Diffusers, loading a Nunchaku checkpoint is as simple as calling from_pretrained(), with no local CUDA compilation required thanks to the kernels package.

---

Record: https://forck.live/items/1452-bringing-nunchaku-4-bit-diffusion-inference-to-diffusers
Catalogue: https://forck.live/llms.txt
Feed: https://forck.live/feed.md