I've noticed this on Qwen3.5 35B before as well, there is a noticeable speed difference between Unsloth's Q4_K_XL and Bartowski's Q4_K_M on the same model, but Gemma 4 seems particularly harsh in this regard: Bartowski gets 38 tk/s, Unsloth gets 28 tk/s... everything else is the same, settings wise.