Qwen3.6-27B KV quantization analysis shows Q8 and Q6 converge with heavy value quantization

A user tested the effect of key-value (KV) cache quantization on the Kullback-Leibler divergence (KLD) of bartowski's Qwen3.6-27B model using llama-perplexity on a 5090 GPU with 50k context.

• Q8 models perform better than Q6 and Q5, with a wider performance gap between Q6 and Q5 than between Q8 and Q6.
• Both Q8 and Q6 suffer steep KLD drops when the value (v) is quantized to q4_0, regardless of key quantization.
• When using q4_0 for values, Q8 and Q6 converge in performance, making (q8_0, q8_0) on Q6 a surprisingly effective compromise.
• Q5 models are more tolerant of value quantization compared to higher precision models.

The analysis suggests that if VRAM constraints force heavy quantization, using full precision KV on Q6 is nearly as effective as Q8, while Q8 remains the preferred choice for maximum coding accuracy when resources allow.