Huawei's Cloud and AI lab officially open-sourced MACaQ, short for Model Agnostic Calibrated Quantization, under an Apache 2.0 license on Thursday.
According to technical documentation published alongside an arXiv preprint, MACaQ can quantize a 122-billion-parameter model in 64 minutes using 32.5 gigabytes of device memory. Published under the huawei-csl GitHub organization, MACaQ targets text, image, video, and audio workloads across major inference engines. According to Emergent, AI Weekly, and AI/TLDR, MACaQ ships with pre-quantized model weights and supports 2-bit, 3-bit, and 4-bit GGUF exports for llama.cpp; it also enables W4A4 and weight-only serving inside vLLM, image and video pipelines in Diffusers, and speech workflows in PyTorch. Active pull requests in the repository document kernel integration for Huawei's Ascend CANN hardware stack. MACaQ treats each linear layer independently by applying per-channel scaling and block-Hadamard rotations, then folding inverse transforms back into weights.
Emergent reported on Thursday that MACaQ quantizes a 122B-parameter model in 64 minutes using only 32.5GB device memory. Authors Muller, Bich, Boretti, Chang, Zhuang, and Cavigelli noted in arXiv paper 2611.10649 that MACaQ matches the best Llama-3-8B perplexity of MR-GPTQ and DuQuant++ with 11 times less quantization time. On Qwen3.8-27B under W4A4 inside vLLM, MACaQ delivered a 2.33x prefill speedup and a 2.58x decode speedup versus BF16, with perplexity shifting from 2.8463 to 2.8994. On Qwen3.6-35B-A3B, its refined Q4_K_XL GGUF achieved lower KL divergence to FP16 than Unsloth's matching 22.36 GB file.
Software analyst SamCodeman noted Thursday that the repository README omits the exact GPU model, batch size, sequence length, and calibration dataset behind the reported vLLM speedups, and clarified that 32.5 GB reflects quantization memory rather than inference VRAM. Another reviewer noted the flagship 122B model lists only Wikitext2 perplexity—5.14 against 5.06 for FP16—without long-context or coding evaluations.
