8B 模型 75MB 投影头对比 Jev：固定工具集快 13 倍

Big Chirpy @loudchirper原推发布于 2026.09.26

在 X 上打开原推

A 75 MB projection head on an 8B model just handed agent loops a 13x speedup over TypeSafe AI's proprietary Jev API. Declaring the closed model dead is the fastest way to break your production pipeline. The architectural trick behind Contrastive Language Models (CLM) on Qwen3-8B is genuine engineering. Jev runs state and choices jointly through a closed cloud model on every step. CLM splits the state head from the action head. If your agent operates on a static set of tools or UI actions, you embed those action candidates once and cache them. In browser agents and gaming loops, that disaggregated caching makes it 4x to 13x faster with zero API bills. When fine-tuned on trajectory data, that tiny open head hits state-of-the-art verifier numbers: 87.6% on Terminal-Bench 2.1 and 81.6% on DeepSWE, where Jev managed roughly 71%. Drop-in primitives like Choice, Noul, and Score mean existing client code swaps right in. The catch is where the open-source victory lap falls apart. Jev is built for zero-shot broad knowledge across massive surface areas. On the Berkeley Function Calling Leaderboard v4, Jev hits 99.2% while CLM-8B slips to 95.2%. CLM also runs on a relative softmax over supplied candidates rather than absolute calibrated scoring, and its representations degrade past an 8K token budget compared to Jev's 64K ceiling. If your agent runs high-frequency decisions over a fixed tool palette, self-hosting a cached 75 MB head is an obvious win. But if your system relies on open-domain common sense across sprawling contexts, closed frontier scale still holds the floor.