Oh I didn't expect this to be on HN haha - but yes for our new benchmarks for Qwen3.5, we devised a slightly different approach for quantization which we plan to roll out to all new models from now on!