TL;DR of the paper, generated automatically.

MACaQ is an open-source 4-bit quantizer for language and diffusion models.

• Method: MXFP4 W4A4 post-training, each layer quantized independently.
• Overhead: only 2.3% over pure MXFP4 on Qwen3-32B in vLLM.
• Speed: matches DuQuant++ and MR-GPTQ's best perplexity in 11x less time.