开源量化方法 MACaQ 在 GGUF 格式上表现超越同类

On November 12, 2026 at 01:32, a discussion appeared in the Reddit community LocalLLaMA about the open-source quantization method MACaQ, where a user published test results run in GGUF format with llama.cpp and claimed that it beats comparable quantization offerings from Unsloth and Bartowski on quality.

In the quantization test on Qwen3-4B (IQ4_XS), the GGUF file produced by MACaQ measured 2.260 GB with a mean KL of 0.02804. Bartowski and Unsloth both produced files of 2.271 GB, with mean KL values of 0.03443 and 0.03663 respectively. The community comparison stated that, at virtually identical file sizes, MACaQ's KL divergence is 23.5% lower than Unsloth and 18.6% lower than Bartowski. In the quantization test on Qwen3.6-35B-A3B (IQ4_XS), MACaQ's GGUF file was 18.86 GB with a mean KL of 0.018696, while Unsloth (UD-IQ4_XS) recorded 17.73 GB and 0.020954, and Bartowski recorded 19.70 GB and 0.021927. The material also mentions a quantization test on Qwen3.8-27B (IQ4_XS), but the accompanying figures are not given in full. No official statement from the MACaQ developers or further progress has been reported, so the matter currently rests at the stage of community-published benchmark results.