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

A new task-aware quantization (TAK) method for Qwen 3.8 27B has achieved 82.81% on reasoning benchmarks, closely approaching the 83.59% performance of BF16 while significantly reducing model size. This TAK quant outperforms Unsloth Dynamic 1.0, 2.0, and 3.0 comparators, demonstrating its effectiveness across various architectures including dense, QAT, and MoE, and models like Gemma 3, Gemma 4, Qwen3.5, and Qwen3.8. Further work can be followed on X or supported via Buy Me a Coffee.

This report details a task-aware quantization method that, unlike earlier versions, consistently beats Unsloth Dynamic comparators across multiple model architectures and Qwen versions.