把 Jev 决策层换成 Mac 本地 1.5GB 小模型

Asteri @Asteri_eth原推发布于 2026.10.01

在 X 上打开原推

Your agent's decision layer can now run on your own Mac. A 1.5 GB local model answers in ~18 ms per call and costs $0 in tokens. Jev itself would bill about $3.02 a month for the same loop. I just wrote a full article on how to set it up: the gate, the model router and the real bill. Here are the 10 steps: 1 → Claude builds. Code checks. A small model decides. The hook allows, asks or denies. 2 → Look at what your safety hook actually ships: every bash command, the folder it runs in, sometimes a line from .env. 600 times a night. 3 → Find the forks with short answers: allow, ask or deny. Those don't need a frontier model. 4 → Put plain code rules first. They catch the obvious cases before any model is called. 5 → Send the narrow questions to a local model. ~18 ms per call, nothing leaves your disk. 6 → Keep fuzzy judgment for Jev or for you. 7 → Pick from 38 open alternatives. Some speak Jev's exact wire format, so old hooks keep working after one base_url change. 8 → Put a router in front, so moving a question between local and Jev is a config change, not a rewrite. 9 → Be honest about accuracy: Von scores 93.5% against Jev's 97.2%. Local is for the questions you never wanted to upload, not for all of them. 10 → Keep state on your machine. Claude builds, code rules filter, the local model decides, the hook enforces. The result: A gate that used to send 600 decisions a night to someone else's API now makes them on your laptop in milliseconds, for $0. Full breakdown below ↓

这条是引用着另一条写的：x.com/asteri_eth/status/2104901781725598103