GitHub radar
Edge0: Run a 35B Model on an Everyday Mac
Edge0's open-source engine keeps only the needed parts of a large model in memory and streams the rest from disk — I review what that means for your API bill.
Edge0 is an open-source engine from the Edge0 team for running large mixture-of-experts models on Macs with Apple Silicon. Instead of loading the whole model into memory, it streams the parts it needs from the SSD and predicts which ones will be needed next, which the authors say speeds up generation by up to 59%. Two preview models ship with it: a 35-billion-parameter version at about 23 GB on disk and roughly 2.9 GB of peak memory, and an 8-billion-parameter version at about 4.2 GB on disk. Both are built on open base models and include adapters that recover quality lost to compression. Only macOS on M1–M4 chips is supported for now.
Why a vibe-coder should care
The money question here is simple: which of your agent's routine calls can run on your own laptop instead of a paid API. Summaries, drafts, sorting inbound messages — if a local model handles them, that part of the token bill goes to zero, and your data never leaves the machine. It is a preview, so I would test it on one repetitive task first and compare the results with the model you pay for.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up Edge0 on my Mac from https://github.com/Edge0-AI/Edge0 — follow the README, start with the 8B model, run it and show me how to ask it something, then tell me whether my Mac can handle the 35B one.
Mac на Apple Silicon (M1–M4): большая модель — около 23 ГБ на диске и около 3 ГБ оперативки в пике, малая — около 4,2 ГБ на диске. Windows и Linux пока не поддерживаются.
Open on GitHub▌ More finds