Analysis of KV quantization effects on Qwen3.6-27B KLD

Study measures Kullback-Leibler divergence across Q8, Q6, and Q5 KV cache quantization for Qwen3.6-27B. Lower KLD indicates better preservation of model outputs; results show increasing quantization reduces quality.