My Qwen3.8-27B task-aware quant reaches 99% of BF16 reasoning performance at 15% of the size.