Which Mac for AI?
A model is a very large book, and the chip has to keep the whole thing open on the desk the entire time it works. Memory is that desk. Cloud AI keeps the book on someone else's desk and posts you the pages, which is why a cheap laptop can use it and cannot run it. Answer five questions and this sizes the desk: the machine to buy, the memory tier that clears the model you named, and the first thing that machine will not be able to open.
Memory is the desk and the model is the book: the chip has to keep the whole book open while it works, so the desk has to be big enough to hold it. Cloud AI keeps the book on someone else's desk and posts you the pages.
Buy this one
Mac mini · M6
24 GB of memory, from $899 (price unverified). It holds a mid-size model (about 27 billion parameters) with the desk still usable.
Your desk
24 GB. Running a mid-size model (about 27 billion parameters) takes about 16.1 GB at 4-bit — that is the model squashed down to a quarter of its full weight — with room for 8,000 words of conversation.
What fits on it
- a small assistant model (about 8 billion parameters) — 5.6 GB — runs with room to spare
- a mid-size model (about 27 billion parameters) — 16.1 GB — runs, but the machine will be busy
- a serious model (about 70 billion parameters) — 41.8 GB — will not open on this desk
- an open frontier model (about 405 billion parameters) — 225.7 GB — will not open on this desk
The first thing this machine cannot open is a serious model (about 70 billion parameters).
One step either way
One step down: Mac mini · M6 · 16 GB — $899 — cheaper, and it nearly works, but the model you asked for will not open on it.
One step up: Mac Studio · M5 Ultra · 512 GB — from $5,499 (unverified — see the note below) — the point where nothing on the list is out of reach any more.
Three things to know before you buy. Memory cannot be added to a Mac later — the chips are soldered to the processor, so the size you choose now is the size for the life of the machine. The 512 GB Mac Studio was announced but does not ship until late October 2026. And these are US list prices read on 2026-09-07, in a year when a memory shortage is moving them — a price marked unverified is one Apple publishes only inside its configurator, so we show the starting price of that machine as a floor rather than invent the upgrade.
Prices from apple.com; model memory from the Local AI Hardware Sizer's arithmetic; the fit rule is ours — a model may take three quarters of the memory before the machine stops being pleasant to use, because macOS and everything else you have open share the same pool.
How to read it: the memory number is the whole decision, and it is the one thing you cannot change afterwards — Apple solders the memory to the processor, so a machine bought at 16 GB is a 16 GB machine forever, however much you later wish otherwise. Storage you can hang off a cable; memory you cannot. That asymmetry is why the honest advice is almost always to spend the upgrade money on memory rather than on a faster chip: a model that does not fit does not run slowly, it does not run at all. Note too that the tool assumes 4-bit weights, which is the ordinary way people run models at home and roughly quarters the size of the book at a modest cost in quality. Run them at full precision and every number here quadruples.
Embed this tool
Paste this snippet into any page to run the picker there.