A Mac mini as your AI server.

1Why a Mac mini

Writing text with a language model is limited by how fast the weights can be read from memory. On a Mac the GPU reads unified memory directly, so a Mac mini with 64 GB can hold models that don't fit on a consumer graphics card, at a few watts when idle. It takes little space and runs macOS, so the same tools work as on a MacBook.

Its limits are memory and bandwidth. The largest Mac mini has 64 GB at 307 GB/s; for 70B-class models at higher precision or the biggest open models, a Mac Studio with up to 512 GB is the step up (Apple). See the best Mac for local LLMs.

2Which Mac mini for which models

Bandwidth sets the speed: the M5 Pro reads memory about 80% faster than the M6 with 24 or 32 GB, so the same model can write up to about 80% faster on it. Each row links to every model that runs well on that Mac, with speeds. Earlier Mac minis with M4 and M4 Pro work the same way: Mac mini M4 Pro.

Agents and long documents need context memory on top of the model. If an AI agent is the main job, pick more memory over a faster chip.

3Keep it awake and reachable

Open System Settings › Energy on the Mac mini (Apple):

• Prevent automatic sleeping when the display is off: on, so the server keeps answering with no screen attached.
• Start up automatically after a power failure: on, so it comes back by itself.
• Wake for network access: on, if you let it sleep.
• Low Power Mode: Apple suggests it for computers that are always on, such as servers, to reduce energy use and fan noise.

To manage it without a screen, turn on Remote Login (SSH) or Screen Sharing in System Settings › General › Sharing (Apple). Set the server app to open at login, so a restart brings everything back.

4Serve a model to your other devices

Each of these runs a model on the Mac mini and gives your other apps an API in the format OpenAI's API uses, which most chat apps, coding tools and agents accept:

With mlx-lm, for example:

pip install mlx-lm
mlx_lm.server --model mlx-community/Qwen3-8B-4bit --host 0.0.0.0 --port 8080

# from another computer on the same network
curl http://<mac-mini-name>.local:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "Hello"}]}'

Ollama unloads a model after 5 minutes without requests by default; set OLLAMA_KEEP_ALIVE to keep it in memory on a server (Ollama). The best apps to run LLMs on a Mac compares these tools in detail.

The AI App Store lists these tools and other apps that run models locally, each with its maker's official download, its licence and where your data goes.

5Running AI agents on it

A lot of Mac minis now run AI agents that work through chat apps, files and the web around the clock. OpenClaw, for example, is an open-source assistant that runs on your own computer, where models, including local ones, are plugins you can swap (OpenClaw). Point such an agent at the local server's address and it uses the model on your Mac mini instead of a cloud API.

Two things matter more for agents than for chat. Context: agents send long prompts with tool descriptions and history, which takes memory on top of the model (see why). And model quality: the biggest model that runs well on your Mac mini will follow instructions and call tools more reliably than a small one.

6Keep it private

• These servers listen only on the Mac itself until you change it. Open them to your home or office network only, and only if you trust the devices on it.
• mlx-lm's own documentation says its server isn't recommended for production because it implements only basic security checks (mlx-lm).
• Don't forward a port to the internet. To reach the Mac mini from outside, use a private network such as a VPN.
• An agent that can run commands or send messages acts with your accounts: give it only the access it needs.

7What it costs, and its idle hours

A home server spends most of its day waiting. MLX Node can use those hours: it runs AI jobs for MLX Network within the limits you set, including how much memory a job may take, and earns MLX Tokens on a free seat, up to 1 MLX a day. Pause it when you need the Mac mini at full speed, or let it start only after the Mac has been idle. Earnings depend on the network and aren't guaranteed: read the risks.

8Questions

Is a Mac mini good for AI?

For running models locally, yes: its GPU reads unified memory directly, so the M5 Pro with 64 GB holds models up to about 85B parameters at 4-bit, at a few watts when idle. For training or the biggest models, a Mac Studio or a GPU server is the better tool.

Which Mac mini should I buy for a local LLM server?

As much memory as you can: memory decides which models fit and can't be upgraded later. The M5 Pro also reads memory about 80% faster than the M6 with 24 or 32 GB, so it writes faster. Compare every configuration at Mac mini M5 Pro and Mac mini M6.

Can a Mac mini run a 70B model?

Only the M5 Pro with 64 GB, at 4-bit and with a short context: the weights take about 39 GB of the 48 GB the GPU can use by default. It writes slowly, a few tokens a second. See Llama 3.3 70B on a Mac.

How do I use the Mac mini's model from my laptop or phone?

Start the server so it listens on your network (for example mlx_lm.server --host 0.0.0.0), then point the app on the other device at http://<mac-mini-name>.local:8080/v1 as an OpenAI-compatible endpoint.

How much power does a Mac mini AI server use?

An M6 Mac mini draws about 4 W at idle and up to 70 W; an M5 Pro 6 W to 145 W. Flat out for a day, that is 1.68 or 3.48 kWh at most (Apple).