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!

Can you describe what is this slightly different approach and why it should work on all models?