What about a Mac mini?

This guide is written for Linux, because it’s the most affordable path, the closest to a real server, and the most faithful to how things actually get deployed. But in real life, plenty of you already have a Mac mini sitting on a desk. Good news: it’s an excellent machine for this purpose, and a large part of what’s written here runs on it without the slightest problem.

What works exactly the same

Most of the journey doesn’t depend on the operating system. On a Mac mini, all of this applies with no adaptation whatsoever:

• Coding agents: Claude Code and OpenCode run natively on macOS. Installing them works the same way.
• Local AI: Ollama has a macOS app, and the models run very well on Apple Silicon. The hybrid cloud + local setup is identical.
• AI agents: the entire chapter (the concept, the setup, agents running in the background, Hermes & OpenClaw) works the same.
• Networking: Tailscale has a macOS app, and Cloudflare Tunnel offers cloudflared for Mac. You expose and reach your machine in exactly the same way.
• Git, GitHub and Docker: identical (Docker Desktop on the Mac side).
• And above all, the whole Working well chapter: scoping a project, memory files, skills, review and audit, maintenance. These are methods, not system commands. They don’t change one bit.

In short: it’s mostly the system layer (the installation and a few admin commands) that differs. Everything else, the heart of the matter, is the same.

What you need to adapt

Four differences to know about, all of them simple:

The 2026 line-up, and what it changes for local AI

Apple refreshed both machines on 25 August 2026, and they have been on sale since 22 September.

The mini gets the M6, Apple’s first chip built on a 2-nanometer process, with a 12-core CPU and GPU and a Dual 16-core Neural Engine, up to twice as powerful as the previous generation according to Apple. Above it, the M5 Pro mini takes the MacBook Pro chip: more cores, up to 64 GB of memory, but a single Neural Engine. Don’t be dazzled by that dual Neural Engine: Ollama, llama.cpp and MLX run language models on the graphics processor, and memory bandwidth sets their speed. On that front, the M5 Pro is 1.8 times faster than the M6.

Frandroid reviewed the Mac mini M6 in its 24 GB version, and its M4 predecessor at launch. On the M6, the Frandroid Lab measured 33.1 tokens per second with a 7-billion-parameter model. The prices shown are for the base versions: €1,049 with 16 GB for the M6 (checked on 30 September 2026), €699 for the M4 at launch. The other machines reviewed, Macs and mini-PCs, are compared in The machines we tested.

The Studio, for its part, doesn’t move to the M6. It keeps the M5 Max and debuts the M5 Ultra, made of two M5 Max chips joined by an interconnect running at over 4.4 TB/s.

What it actually costs

Starting prices checked on apple.com/fr (France) on 1 October 2026:

The machine links lead to Quelle IA (in French), which lists the models each configuration can run.

Memory costs €27.50 per gigabyte: on the M6 mini, going from 16 to 24 GB costs €220 at equal storage, and the Apple configurations I checked follow the same rate.

Your targets therefore come down to this, in the configurations offered by Apple (SSD included, which weighs on the price per GB):

Look at memory AND throughput, not one or the other

The usual reflex is to count only RAM, since it decides whether a model fits. True, and insufficient.

Once the model is loaded, what sets generation speed is memory bandwidth. Every token produced forces the machine to reread all of the model’s weights, so speed tops out at bandwidth divided by model size. Apple’s ladder climbs clearly at each step:

Which translates very concretely into the model’s typing speed:

These are theoretical ceilings, reality runs rather between sixty and eighty percent of these values. But the order of magnitude is enough to decide: a 70B doesn’t even fit in an M6 mini, it types slowly on a 64 GB M5 Pro where it brushes the memory limit, and it becomes usable on a well-equipped M5 Max Studio. For a specific model, each machine page on Quelle IA (in French) gives an estimated speed, labelled as such.

Three targets, depending on what you do with it

Small home workshop: Mac mini M6 32 GB, €2,039 with 1 TB. Tiny power draw, silent, and enough memory for a quantized thirty-billion-parameter model. It’s the most sensible entry into local AI on a Mac. Avoid 16 GB on a new machine dedicated to this.

Maximum capacity on a mini: Mac mini M5 Pro 64 GB, €3,649 with 1 TB. This is the threshold that changes what you can load, since you leave the 30B class. A 70B in 4-bit brushes the limit there: Quelle IA counts about 47 GB usable by the GPU on a 64 GB Mac, for roughly 48 GB needed. Also know that you pay nearly 1.8 times the price of the 32 GB M6 mini.

Workshop built to last: Mac Studio M5 Max. The only one of the three that goes up to 128 GB, and the one whose memory bandwidth (614 GB/s with the 40-core GPU, required above 36 GB) makes the biggest difference once the model is loaded. With 64 GB and 1 TB, it costs €4,429. Apple doesn’t publish the price of each memory tier, so check it yourself in the configurator before deciding.

What it draws, and why that is an argument

On a machine that stays on permanently, power draw stops being a detail: it becomes a line on a bill and a source of heat in the room.

Apple publishes the maximum continuous power of its machines: 155 W for the Mac mini, 480 W for the Mac Studio. These are ceilings, rarely reached, and an idle machine draws far less. For comparison, Nvidia lists 600 W for a single RTX Pro 6000 card, not counting the PC around it.

Here is what a few average power levels cost, running all year:

Calculated at twenty-five cents per kilowatt-hour, the order of magnitude of the French rate.

So, Linux or Mac mini?

Both are excellent choices, and it mostly comes down to what you have on hand. A Linux mini-PC is cheaper at equal memory, more “server”-like, and hugs the deployment environment as closely as possible. A Mac mini, you may already own one, its unified memory is an asset for local AI, and macOS is probably more familiar to you. Either way, the real subject of this guide stays the same: building a workshop where you describe a project and the machine builds it with you.