Drex 2.0 对 Jev:3B 模型在决策游戏里小胜
256 局对战里 Drex 2.0 决策指数 61.0 对 Jev 58.0,胜率 60.9%,主打 agent 路由和工具选择这类小决策。
Drex 2.0 just gave Jev a serious problem with only 3B active parameters. two models go head-to-head in a live game. same board, constant decisions, and every bad move is right there on screen. → Drex 2.0: 61.0 Decision Index → Jev: 58.0 → 60.9% head-to-head score across 256 games → up to 38% faster on 32K–128K inputs the crazy part? @NaceAI built this to make decisions without dragging a massive model through every single step. $0.06 per 1M input tokens makes that pretty interesting. think agent routing, tool selection, and hundreds of tiny decisions happening before the expensive model even wakes up. the game is fun to watch. the real win is making AI agents stop paying full price just to decide what happens next.
这条是引用着另一条写的:x.com/naceai/status/2109703175326601579