GGUF Quants Arena Tops MMLU Subset Scores

The main signal here is that quantization choice matters, but not always in the way people expect: the highest-scoring Q5 quant barely edges the Q4 quant, and a few alternative 27B variants stay close enough to make deployment tradeoffs more about memory and speed than raw accuracy.

• –Qwen3.5-27B-UD-Q5_K_XL and Q4_K_XL are effectively tied, so the cheaper quant may be the better default if runtime constraints matter
• –The distilled 27B reasoning model lands close behind, suggesting distillation can preserve benchmark strength surprisingly well
• –Smaller and more aggressive quants fall off steadily, which is consistent with MMLU punishing over-compression more than casual chat workloads do
• –The post is most useful as a practical llama.cpp tuning reference, not as a model-launch announcement
• –The single error noted on gemma-4-31B-it hints that benchmark hygiene still matters as much as the score table itself