Mac mini for local AI: which configuration to buy

A Mac mini is the usual suggestion for a private AI machine on a desk: a small box with Apple silicon's unified memory, and a fan. The question is which one. Memory decides what you can run at all; the chip decides how fast. Here is what each configuration can hold, based on model sizes and speeds we measured — and, clearly marked, what we could not measure.

The current line-up

From Apple's Mac mini specifications, checked on October 12, 2026. Storage starts at 256 GB on the 16 GB model and 512 GB on the others. Prices vary by country and change; see Apple's store for yours.

Memory cannot be added later, so this is the decision to get right.

What we measured, and on what

Not on a Mac mini. Our numbers come from a MacBook Air with an M5 chip — 10-core GPU, 24 GB, 153 GB/s by Apple's figure — which is one chip generation behind the mini's M6 and has no fan. Read them as a likely floor for the M6 mini, not as its score.

Memory is the model file plus its working area for a 4,096-token context. Full tables: speed, memory.

Two differences should favour the mini, and we have measured neither. The M6 has two more GPU cores, and in its 24 and 32 GB configurations more memory bandwidth than our test machine. And it has a fan: the Air lost about 41% of its generation speed after two minutes of continuous work, which is a cooling limit, not a chip limit.

16 GB: every small and mid-size model

A model needs its file plus a working area in memory, and half of the machine is a comfortable budget for it. Half of 16 GB is 8 GB, which covers everything up to 9B in our test — the largest, Ornith 1.5 9B, needs 6.1 GB. It also covers all 22 models in the privateSLM catalog; the largest file there is 4.9 GB.

What it does not hold comfortably is the 20B class: gpt-oss-20b needs 12 GB before you open anything else. Note too that the 16 GB model starts at 256 GB of storage; the 12 model files in our test add up to 40 GB.

24 GB: the 20B class

This is the configuration we can speak for, because our test machine has the same amount. A 20-billion-parameter model with a 32,000-token context took 13.4 GB and ran at 48 tokens a second — faster than the 7B models, because it is a mixture-of-experts design that reads only part of itself for each token. If you are buying a mini mainly to run local models, this is the one we would start from: it is the smallest configuration that runs that class comfortably, and it is also where the M6's faster memory begins.

32 GB: headroom, by arithmetic

We have not run anything that needs it. By the 0.6 GB-per-billion-parameters rule our files follow, a 27B model is about 16 GB — half of this machine. Buy it if you already know you want models of that size, or if the mini will be doing other memory-hungry work at the same time.

M5 Pro, 48 or 64 GB: larger models, and faster ones

Again arithmetic: a 70B model at 4 bits is around 42 GB, which is workable in 64 GB and too tight for 48. The other reason to look at the M5 Pro is the last column of the table above. Generating text is mostly a matter of reading the model's weights from memory over and over, so memory bandwidth is the specification most often tied to speed on large models, and the M5 Pro has nearly twice the M6's. We have not measured by how much that helps, and would not guess.

What we would buy

• To try local AI and keep the option open: M6 with 24 GB.
• Only small and mid-size models, and the budget is tight: M6 with 16 GB — with more than 256 GB of storage if you can.
• You already know you want 27B-class models: M6 with 32 GB.
• Larger than that, or speed on big models matters: M5 Pro with 48 or 64 GB.

What this page does not know

• How fast any Mac mini actually is. Every speed here was measured on an M5 MacBook Air.
• How models larger than 20B behave. Everything above that line is arithmetic from file sizes.
• Prices. They differ by country and by the week.
• Answer quality. A bigger model is not automatically the better one for your work.

The measurements behind the numbers, with code and raw data: bench.c, run_all.sh and the raw results.jsonl (55 runs).

Once it is on your desk

privateSLM is free on the Mac App Store and runs on the mini as it does on a MacBook or an iPhone: on a Mac with Apple Intelligence it answers straight away with the built-in model, and the catalog adds open models that download once and then work offline. Nothing you type leaves the machine.