Also: a heavily quantised recent model beats its full precision older versions.

Qwen 3.8 27B Q3 readily beats Qwen 3.6 27B FP16. Difference in hardware requirements and token rate is of course massive.