The larger 3.5 quants are actually pretty close to the full-blown 397B model's performance, at least looking at the numbers. Qwen 3.5 seems more tolerant of quantization than most.