what quants are people using for the big Qwen3? between the new version and the old, I can't get it to output commas during narrative sequences. Is it wrong to use ChatML or something? I've used the recommended settings, neutral samplers, and a variety of more or less permissive settings than the recommended. Makes no difference what I do. Currently using the instruct at q3_s, but I've also used the q3_XL from unsloth or whatever. I'm wondering if it's quant degradation, or if the unsloth guys fucked up the quant I downloaded (I know they've had issues before). pls send help