RCreddit.com·
暂不在当前实时榜单
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.
为什么是这条This report details a task-aware quantization method that, unlike earlier versions, consistently beats Unsloth Dynamic comparators across multiple model architectures and Qwen versions.
时间与来源
时间显示为 UTC
显示时区:UTC
本地时区尚不可用,暂时显示 UTC。
收录当时偏移:UTC+02026年9月8日 08:00 UTC
- 收录
- 2026年9月8日 08:00
- 来源类型
- 开发者社区
正文
本站未收录正文。
前往源站阅读 →