The Cheapest Mac That Runs Serious Local AI

People ask us this constantly, and they almost always ask it the wrong way. They ask which Mac is cheapest. That question has a clean answer and it is the wrong answer, because the cheapest Mac in the lineup will run a model badly and teach you that local AI is slow and frustrating, which is not true. The right question is buried one layer down: what is the cheapest Mac that runs serious local AI well. That one has an answer we trust, and it almost never points at the headline-cheapest machine.

So here is how we actually reason about it, and what we tell people who want a starting point without overspending.

Reason by RAM, not by price tag

The single lever that decides whether a Mac is good at local AI is memory. Not the chip generation, not the core count, not the model year on the box. Memory.

This is because of how Apple Silicon is built. The chip uses unified memory, which means the CPU and the GPU share one fast pool instead of copying data back and forth between separate banks. A language model has to sit entirely in that pool to run, and the model is large. The amount of memory you have is the size of the model you can hold, and the size of the model is most of what determines whether the output is good enough for real work.

So when someone shows us two Macs and asks which is the better buy, the first thing we do is ignore the prices and look at the RAM. A cheaper machine with more memory beats a pricier machine with less, almost every time, for this specific job. That is the whole trick, and once you see it you cannot unsee it.

The numbers that actually matter

We work from two thresholds, both grounded in what the runtimes ask for.

16GB is the floor for serious work. The Atlas notes on both Ollama and MLX land on 16GB as the recommended amount, and that matches our experience: a useful coding model runs on 16GB with enough headroom to keep an editor and a browser open beside it. You can technically run a small model on 8GB, but you will spend the whole time fighting for room, and the moment you open anything else the machine starts swapping. We do not recommend 8GB for anyone who wants to do this for real.

32GB is where it gets comfortable. The MLX note calls for 32GB or more once you want larger models, and that is the line we would draw too. With 32GB you can run a bigger, smarter model and still have your normal work open at the same time, without the careful memory accounting that 16GB sometimes demands. If your budget can stretch one notch, this is the notch to stretch for.

Past 32GB you are into specialist territory: very large models, multiple models loaded at once, heavy batch jobs. Most people, including us most days, do not need it. Spend the money on getting to 32GB cleanly before you spend it anywhere else.

Why a used Mac is usually the value play

Here is the part that surprises people. The cheapest sensible entry into serious local AI is often not a new machine at all. It is a previous-generation Apple Silicon Mac with the right amount of memory, bought used.

The reason is that the thing you are paying for, unified memory, does not really go stale. An older Apple Silicon chip with 32GB will hold and run the same model a newer chip with 16GB cannot. It may run it a little slower, but slower-and-it-fits beats faster-and-it-does-not-fit every single time, because the second machine simply cannot do the job. When you reason by RAM, the used market suddenly looks very different: you are hunting for memory, and memory on a one or two year old machine costs much less than memory on this year’s.

We will not quote you model years or chip names as gospel here, because the lineup shifts and a fact we print today goes stale by next quarter. [CONFIRM: current entry-level Apple Silicon models and their generations]. What we will say is the rule that does not go stale: find the oldest Apple Silicon Mac you are comfortable owning, then buy the most memory you can on it. [CONFIRM: typical used price for a 32GB previous-generation Mac in your region].

A few honest cautions on the used path. Check the memory configuration before you buy, because you cannot upgrade it later: Apple Silicon memory is soldered, so the number on the listing is the number you keep forever. Confirm it is Apple Silicon and not an older Intel Mac, which will not run any of this well. And check battery health if it is a laptop, because that is the one part of a used machine that genuinely wears out.

What we run, and what we would buy today

Our studio machines are Apple Silicon, chosen for exactly the reason above: unified memory is genuinely good at this work, and it is hard to beat for the money on a single desk. We run Ollama as the everyday engine and reach for MLX when a long job makes speed the bottleneck, and both of those live comfortably inside the memory thresholds we just described.

If we were starting someone from zero today, on the smallest budget that still works, we would not point them at the cheapest new Mac. We would point them at a used Apple Silicon machine with 32GB if the budget allows and 16GB if it does not, and we would tell them to spend nothing extra on chip speed until the memory is sorted. [CONFIRM: the specific used configuration the studio currently recommends as a starting point].

The takeaway

The cheapest Mac that runs serious local AI is the cheapest Mac that has enough memory, and that is a different machine from the cheapest Mac. Reason by RAM. Treat 16GB as the floor and 32GB as comfortable. Look hard at the used and previous-generation market, because that is where memory gets cheap. And ignore the leaderboard of chips entirely until the gigabytes are right, because on this job the gigabytes are the game.

Buy the memory. The rest sorts itself out. Curious about these things. You should be too.