2026-11-11 · America/Los_Angeles · 社区动态 · #5

Open-source method (MACaQ) beats Unsloth GGUFs!

Some results for those using GGUFs and llama.cpp: GGUF quality: MACaQ vs. Unsloth and Bartowski Qwen3-4B (IQ4_XS) Quantization GGUF size Mean KL ↓ MACaQ 2.260 GB 0.02804 Bartowski 2.271 GB 0.03443 Unsloth 2.271 GB 0.03663 23.5% lower KL divergence than Unsloth and 18.6% lower than Bartowski, at virtually identical file sizes. Qwen3.6-35B-A3B (IQ4_XS) Quantization GGUF size Mean KL ↓ MACaQ 18.86 GB 0.018696 Unsloth (UD-IQ4_XS) 17.73 GB 0.020954 Bartowski 19.70 GB 0.021927 Qwen3.8-27B (IQ4_XS) Quantization GGUF size Mean KL ↓ MACaQ 14.96 GB 0.036784 Bartowski 15.48 GB 0.104170 ~ 65% lower KL…

热度 47.9 / 100；排名与评分保留该期记录。