智能模型关掉推理更省
Turn reasoning off in Hermes
要点
- 模型够聪明时，Hermes 里约百分之九十九的任务应关掉推理。
- 带工具的套件里推理会把 token 账单迅速打爆。
- 低延迟能保住代理循环，等待过长会把整条工作流拖死。
- 作者倾向用 GPT-6 Astra 的即时或低推理档作为 Hermes 默认。
原帖开头
This is actually so underrated for agentic work Hermes is probably best with reasoning ~off for 99% of tasks if the model is smart enough --token cost can blow up so fast in harnesses like that, and low latency is a huge boost Gpt-6-astra instant/low will be the best hermes …
