Hugging Face AI Updates: July 24, 2026

1. Hugging Face Adds Native Nunchaku 4-bit Diffusion Inference to Diffusers

Hugging Face. Hugging Face published a post on July 23 detailing native support for Nunchaku 4-bit checkpoints in Diffusers through an integration called Nunchaku Lite, removing the need for a separate inference engine or local compilation. Nunchaku is built on SVDQuant, a quantization method that runs diffusion transformers with 4-bit weights and activations by moving activation outliers into weights and handling the hardest parts of weight matrices with a small 16-bit low-rank branch. The integration reports roughly a 30 percent speedup and up to 50 percent lower peak VRAM while loading models through standard from_pretrained() calls and remaining compatible with LoRA, offloading, and torch.compile. Source