MACaQ quantizes 122B model to W4A4 in 64 min on 32.5GB GPU

MACaQ, a new post-training quantization method, compresses a 122-billion-parameter model to MXFP4 W4A4 in 64 minutes using 32.5GB of GPU RAM, according to a paper on arXiv by Lorenz K. Muller, Philippe Bich, Chiara Boretti, Hyun-Min Chang, Jiawei Zhuang and Lukas Cavigelli. The authors report matching the perplexity of DuQuant++ and MR-GPTQ on Llama-3-8B in "11 times less quantization time," while adding only "2.3% end-to-end overhead over pure MXFP4 on Qwen3-32B in vLLM."

The speed comes from treating every linear layer independently, avoiding the cross-layer propagation of basis transforms that prior W4A4 methods required. "Retaining these transforms online avoids cross-layer dependencies and enables a simple streaming implementation of the quantization process, in which only one layer must reside in memory at a time," the paper states.

The method pairs a per-channel scale and a block-Hadamard transform on activations with a new "Flip-Refinement" procedure, described as a Hessian-aware coordinate-descent that greedily moves individual FP4 codes to adjacent representable values. The authors also apply MACaQ without architecture-specific modifications to an image diffusion model, which they offer as evidence that the layer-local design generalizes beyond language models.

The abstract publishes no downstream task accuracy numbers and does not identify which 122B model was used for the headline 64-minute result.