Try out this "high" reasoning mode for 27B (tested on VLLM)

After a lot of tweaking, I have come to the conclusion that 27B lacks a reasoning mode that is between low and xhigh. The "medium" mode isn't actually medium, it erases the explicit instructions to the model. When medium is enabled, the model acts very differently - to me it looks like it regresses

中文全文翻译正在补齐，可先查看下方英文原文。

After a lot of tweaking, I have come to the conclusion that 27B lacks a reasoning mode that is between low and xhigh. The "medium" mode isn't actually medium, it erases the explicit instructions to the model. When medium is enabled, the model acts very differently - to me it looks like it regresses