Issue 2026-11-11 · Industry · 开源 · 研究
Huawei open-sources MACaQ quantizer for local inference

Huawei released MACaQ, an internal quantizer supporting llama.cpp, vLLM, Diffusers and PyTorch across text, image, video and audio models. Huawei says it can quantize a 122B-parameter model in 64 minutes with 32.5 GB of device memory, matching or beating MR-GPTQ and DuQuant++ perplexity with far less quantization time.
Read original ↗