We all know modern "intelligent" Quantization that uses an imatrix to make a Q4_K_XL model to feel like Q6_K. But here is what i notice: While this works well on most English tasks, the effect can be reversed on other languages or niche tasks. The reason is quite simple and you will find out quickly
