Follow-up to my KV cache quantization post: six weight quants (Q3_K_XL through Q5_K_M) swept across the full KV cache grid, to see where weight quantization stops being worth the VRAM it saves.
I measure how much KV cache quantization (F16/Q8/Q4) hurts Qwen3.8-27B output quality via KL divergence, and how that compares to the impact of quantizing the model weights themselves.
