Huawei CSL (huawei-csl) · Model quantisation toolkit

MACaQ

MACaQ is an Apache 2.0 licensed, model-agnostic calibrated quantisation toolkit for producing GGUF, vLLM, Diffusers and PyTorch checkpoints for text, image, video and audio models. The page reports quantising a 122B-parameter model in 64 minutes using 32.5 GB of device memory, and lower perplexity than MR-GPTQ and DuQuant++. For Qwen3.6-35B-A3B, its MACaQ-R Q4_K_XL GGUF (22.36 GB) shows a KL divergence to FP16 of 0.013367, against 0.015731 for the Unsloth UD-Q4_K_XL file of the same size.