I use Qwen 3.6 35B-A3B constantly, but I don’t see the type of behavior you mentioned. I’m using Unsloth’s Q8_K_XL quant.