Quantization Ranking Battle

Compared GGUF quantized models in a llama.cpp environment based on MMLU subset (DEV+TEST).

• Settings: ctx 8192, seed 42, fa on
• The top tier showed almost no difference between Qwen3.5-27B-UD-Q5_K_XL.gguf 87.33% and Qwen3.5-27B-UD-Q4_K_XL.gguf 87.25%.
• Next was Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled.i1-Q4_K_M.gguf 87.02%.
• Qwen3-Coder-Next-UD-Q4_K_XL.gguf followed at 84.38%, and Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf recorded 83.25%.
• The smaller Qwen3.5-9B-UD-Q8_K_XL.gguf scored 78.81%, and gemma-4-31B-it-UD-Q4_K_XL.gguf scored 78.36%, with errors=1 noted.
• The largest model, Qwen3.5-397B-A17B-UD-IQ2_XXS-00001-of-00004.gguf, was relatively low at 65.80%.

This is a benchmark comparison that shows at a glance the MMLU performance differences across various GGUF quantization combinations under the same conditions.