300 美元二手 Mac mini 本地跑 Jev 等工具替代 800 美元订阅

starmex @starmexxx原推发布于 2026.10.11

在 X 上打开原推

this is pure f*cking treasure. a used $300 mac mini from 2020 quietly killed my $800 a month ai bill, and it runs on $5 of electricity. took me three weekends of testing, so you don't have to. the subscriptions i was bleeding every month: cursor ultra .............. $200 claude code max 20x ....... $200 google ai ultra 20x ....... $200 chatgpt pro ............... $200 total ..................... $800 what i pay now: used mac mini m1, 16gb .... $300 once, ebay electricity ............... $5/mo subscriptions ............. $0 here's what runs on a 6-year-old mac, all local: ollama .................... ollama com one command and gpt-oss-20b, qwen or deepseek is running on the mac. gpt-oss is openai's own open-weight model, the same company charging $200 a month for chatgpt pro gave this one away for free lm studio ................. lmstudio ai a chatgpt-style desktop app for local models. search hugging face, click download, start chatting. no terminal at all mlx-lm .................... apple's own framework apple's open-source library built for apple silicon. runs models straight on unified memory, the fastest way to squeeze every token out of an m1 fast-browser-use .......... qwen3.5-9b on mlx a browser agent that clicks through real websites with zero cloud calls, fits in about 7gb, and plugs into claude code and codex as a skill rizzo-flow ................ 4.4gb the local clone of typesafe's jev. yes / no / pick one / score with a probability on every answer, and the runtime is 11mb on a mac exo ....................... a second mac mini links several macs into one cluster and splits bigger models between them. two $300 mac minis on a shelf start acting like one much bigger machine what nobody tells you until you actually do it: 16gb is the real limit, not the model card. gpt-oss-20b eats almost all of it, so close chrome before you load it or macos starts swapping and everything crawls -> always grab the mlx build in lm studio, on apple silicon it's noticeably faster than the generic one -> set ollama to keep the model loaded, otherwise every first reply waits while it reloads from disk -> run the mac mini headless in a closet and hit it from your phone over tailscale, it feels exactly like the chatgpt app models and repos in the comments. save it before your $800 auto-renews

这条是引用着另一条写的：x.com/starmexxx/status/2108159970017488953