“On 16GB you cannot run a full agentic coding model — the good ones (30B+) simply don’t fit. Anyone telling you otherwise is quietly on a 64GB machine.”
Two months later, guess who’s telling you otherwise? Me!!
I ran a super-huge LLM on a 16GB MacBook Pro M2 from 2022. GLM-5.3 — 744 billion parameters, which everyone calls “the most capable open-weights model for coding”. On disk, it’s 419GB. That’s 26× my MacBook’s RAM.
And it ran! I’d say crawled instead.
It solved a puzzle.
It wrote me a game you can play right now.
It cracked a math riddle in ~9 hours — then spent four more days making sure.
And it told me its name is Claude! More on this in #4.
For the past few weeks I have honestly been considering getting a Mac mini as a home server to run a local LLM, so this story was another push for me. Loved how you covered everything from beginning to end along with the FAQ - 18 words/min might be…
Running a 744B model on just 16GB is impressive, even if it's painfully slow. The SSD swapping issue was especially interesting. It shows that getting a massive model to run is one thing, but making it practical is another.
