Huawei Computing Systems Lab (CSL), the company's research group in Switzerland, has open-sourced MACaQ, a quantization toolkit it says it had been using internally for several months. The huawei-csl/MACaQ repository went public on GitHub on November 10, 2026 under the Apache-2.0 licence, accompanied by a project page and the paper "MACaQ: Fast, Model Agnostic Calibrated Quantization for Low-Precision Weights and Activations" (arXiv 2611.10649). The authors are Lorenz K. Muller, Philippe Bich, Chiara Boretti, Hyun-Min Chang, Jiawei Zhuang and Lukas Cavigelli.
The name stands for Model Agnostic Calibrated Quantization, and the method's selling point is that it never reaches beyond a single linear layer. Many recent low-bit methods rotate or rescale activations and then fold the inverse transforms into neighbouring layers, which ties the recipe to a specific architecture. MACaQ instead applies a per-channel scale and a small block-Hadamard transform to each layer's activations at inference time and absorbs the matching inverse into that layer's weights. Because no other layer is touched, the model can be calibrated and quantized as a stream, with only one layer held in memory at a time.
On top of that, the paper introduces Flip-Refinement, a Hessian-aware coordinate-descent step that nudges individual FP4 codes to adjacent representable values instead of relying on GPTQ-style updates. The core target format is MXFP4 with 4-bit weights and 4-bit activations (W4A4). According to the authors, running the activation transforms online adds only 2.3% end-to-end overhead over plain MXFP4 on Qwen3-32B in vLLM.
The repository is organised by backend. A llama.cpp path produces 2-, 3- and 4-bit GGUF files for dense and mixture-of-experts text models from the Qwen3 family, OLMoE and others, and a vLLM path serves W4A4 or weight-only checkpoints. A Diffusers path covers image, video and audio generators such as FLUX.2-klein-4B and Wan2.2, and a PyTorch path handles speech models. An Ascend section points to pending CANN kernel pull requests, and the repository says full-model execution on Huawei's own accelerators still needs integration with an Ascend inference runtime.
A separate mode, MACaQ-R, refines the weight codes inside an existing GGUF while keeping its native formats and block scales, so the output runs on a standard llama.cpp build. For Qwen3.6-35B-A3B at an identical 22.36 GB file size, Huawei CSL reports a KL divergence to FP16 of 0.013367 for its Q4_K_XL file, against 0.015731 for Unsloth's widely used UD-Q4_K_XL quant.
The paper's headline numbers are also the authors' own. They report lower perplexity than DuQuant++ and MR-GPTQ, a match for those methods' best Llama-3-8B perplexity in 11 times less quantization time, and a 122-billion-parameter Qwen3.5 model quantized in 64 minutes with 32.5 GB of GPU memory. That model's WikiText-2 perplexity moved from 5.06 to 5.14, and the 74 GB MXFP4 result fits on a single 96 GB GPU that could not hold the original weights.
The release drew attention in the local-inference community quickly. A Hacker News thread titled "Huawei releases MACaQ, its internal quantizer for local inference" passed 100 points within a day, and commenters pointed to users already producing their own MACaQ GGUFs. Independent benchmarks have not yet appeared, so the quality and speed claims rest on Huawei CSL's paper, project page and repository for now. Results will also vary with calibration data, model family and hardware, which the authors acknowledge by asking for representative calibration text in every workflow.