Whether you can run a full AI model at home depends entirely on what “full” means.
A 7B model? Probably a laptop with as little as 32GB of RAM — a $1500 MacBook Air. Bigger models, the 70B class, and you probably need 128GB. Then you get to the 1T models, and you’re looking at commercial hardware — though people are doing it with $40K worth of Mac Studios linked together, at about a third of the speed of a cloud service.
That $40K isn’t folklore. Apple’s M3 Ultra Mac Studio can be configured with up to 512GB of unified memory — the most ever put in a personal computer, by Apple’s own account — and Apple itself pitches it as able to run models “with over 600 billion parameters entirely in memory.” In December, Creative Strategies chained four of them together — two with 256GB and two with 512GB, “roughly $39,596 before tax” — over the RDMA-over-Thunderbolt 5 link Apple switched on in macOS 26.2, and ran Kimi K2 Thinking, a one-trillion-parameter model, at about 25 tokens per second. A trillion-parameter machine on a desk, for the price of a decent car, slower than the cloud and answerable to nobody.
I’m very keen to know what people find. I’ve heard really positive spin and negative spin on the home machines, and I don’t know anyone with a couple of weeks under their belt. I want to hear from someone who has lived with one.
Also worth considering, and pulling the other way: Moonshot — their V2.5, Kimi K2.5 to give it its proper name. It’s a fraction of the cost of Claude, so perhaps like having a home machine price-wise, but with huge throughput. It’s also, as it happens, a trillion-parameter model itself — 1T total, 32B active on any given token — with open weights under a modified MIT license. Which means the same class of model you’d need $40K of Macs to host at home is on offer as a rented service, at a fraction of Claude’s rate. Whether that makes the home machine a hobby or a tool is the question I can’t answer yet.
I need to probe further.
