Drex 与 Jev 对比：先做便宜决策再调大模型
作者把路由层拆出来：先用小决策模型排序动作、12ms 分流简单分支，只在需要深推理时升级；Drex 决策指数 51.73 略胜 Jev 1.13.0 的 51.67，速度快 3 倍。
Drex vs Jev is one of the first times a tiny decision model actually made the routing layer feel worth building. in my workflow, the big win is making the cheap decision before the expensive model starts thinking. less wasted calls, faster retries, tighter loops. → rank actions before spending tokens → route simple branches in 12ms → keep 32k context for state + tool options → escalate only when deeper reasoning is needed drex scored 51.73 on the decision index vs 51.67 for jev 1.13.0, while running 3x faster. built by http://Nace.AI, it’s basically a small model for routing, tool selection and reranking instead of another chatbot. i’d use drex for decisions and heavier models for execution. first 10,000 devs also get 250M tokens free.