Hqq Quantization

This gives Claude the ability to quantize neural network weights down to extremely low bit depths (8/4/3/2/1-bit) without needing calibration data, which is the main selling point. It's built on Half-Quadratic Quantization and works with HuggingFace Transformers and vLLM out of the box. You'd reach for this when you want to compress a model quickly without hunting down calibration datasets, or when you're experimenting with aggressive quantization levels that other methods don't support well. The PEFT compatibility means you can fine-tune these quantized models with LoRA, which is handy for keeping memory usage low during training. It's clearly faster than GPTQ or AWQ for the quantization step itself, though inference performance will depend on which backend you pick.