Mac mini M4 Pro for Local AI: Complete Setup Guide (2026)

The Mac mini M4 Pro is the most overlooked local AI machine in 2026. At $1,299 on Amazon (frequently $100 off MSRP), you get a complete, silent, low-power system with 24 GB of unified memory and 273 GB/s memory bandwidth. That bandwidth figure matters: it is what determines how fast tokens generate during LLM inference. The Mac mini M4 Pro runs 13B models at approximately 50 tokens/second and 32B models at approximately 20–25 tokens/second via MLX — something no discrete GPU at this price can match, because no $1,300 discrete GPU has more than 16 GB of VRAM. This guide covers everything: which model to buy, how to set up MLX and Ollama, which LLMs to run, and how to use it as a private AI server.

Which Mac mini M4 Pro to Buy

The Mac mini M4 Pro comes in two memory configurations that matter for AI work. The 24 GB configuration ($1,399 MSRP, ~$1,299 on Amazon) is the right starting point for most users. It runs 7B models at full speed, 13B models comfortably, and 32B models at Q4 quantization. The 48 GB configuration ($1,799) is the upgrade path if you want to run 32B models at Q8 (near-lossless quality) or 70B models at Q4. There is also a 64 GB configuration ($1,999) but at that price the Mac Studio M4 Max 36GB ($1,999, 410 GB/s bandwidth) becomes the better buy for 32B models — it has significantly higher memory bandwidth and a more powerful GPU. Avoid the base M4 Mac mini (not M4 Pro) for serious AI work: it has only 120 GB/s memory bandwidth, less than half the M4 Pro's 273 GB/s.

Memory Configuration Guide

Use this table to choose the right configuration based on the models you want to run. All token speeds are approximate MLX inference speeds on the M4 Pro chip.

Buy Links

The Mac mini M4 Pro is available directly from Apple and via Amazon. Amazon frequently runs $100 off promotions. Buy Mac mini M4 Pro 24GB on Amazon (~$1,299). Buy Mac mini M4 Pro 48GB on Amazon (~$1,799). You will also need a display, keyboard, and mouse if you don't already have them — the Mac mini ships without peripherals.

Step 1: Install MLX (Apple's AI Framework)

MLX is Apple's machine learning framework, optimized specifically for Apple Silicon. It uses the Neural Engine and unified memory architecture to deliver 3–4x faster inference than Ollama/llama.cpp on the same hardware. Install it first — it is the primary inference engine you will use on the Mac mini M4 Pro.

# Install MLX and the LLM server package
pip install mlx-lm

# Verify installation
python -c "import mlx; print('MLX version:', mlx.__version__)"

# Run a quick inference test with a small model
mlx_lm.generate \
  --model mlx-community/Qwen3.5-9B-Instruct-4bit \
  --prompt "What is private AI?" \
  --max-tokens 200

Step 2: Install Ollama (for OpenAI-Compatible API)

Ollama provides an OpenAI-compatible REST API that lets your existing tools (Open WebUI, LangChain, LlamaIndex, Continue.dev) connect to your local model without code changes. Install it alongside MLX — use MLX for direct inference speed, Ollama for API compatibility with other tools.

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull recommended models for Mac mini M4 Pro 24GB
ollama pull qwen3.5:9b          # 14B model, excellent quality, fits in 24GB
ollama pull llama3.2:3b          # Fast 3B model for quick tasks
ollama pull deepseek-r1:14b      # Best reasoning model at this size

# Start the Ollama server (runs on port 11434)
ollama serve

# Test the API
curl http://localhost:11434/api/generate -d '{
  "model": "qwen3.5:9b",
  "prompt": "Explain unified memory in one sentence."
}'

Step 3: Run 32B Models with MLX

The Mac mini M4 Pro's key advantage is running 32B models — something no $1,300 discrete GPU can do. Use MLX for best performance on large models. The mlx-community Hugging Face organization maintains pre-quantized MLX versions of all major models.

# Run Qwen2.5 32B at Q4 quantization (fits in 24GB, ~22 tok/s)
python -m mlx_lm.generate \
  --model mlx-community/Qwen3.5-27B-Instruct-4bit \
  --prompt "Write a Python function to encrypt a file with AES-256" \
  --max-tokens 500

# Run as a server (OpenAI-compatible API on port 8080)
python -m mlx_lm.server \
  --model mlx-community/Qwen3.5-27B-Instruct-4bit \
  --host 0.0.0.0 \
  --port 8080

# On 48GB config: run at Q8 for near-lossless quality
python -m mlx_lm.server \
  --model mlx-community/Qwen3.5-27B-Instruct-8bit \
  --host 0.0.0.0 \
  --port 8080

Step 4: Set Up Open WebUI (Private ChatGPT Interface)

Open WebUI gives you a full ChatGPT-style interface connected to your local models. Install it with Docker and point it at your Ollama instance. Your conversations never leave the Mac mini.

# Install Docker Desktop for Mac from docker.com, then:
docker run -d \
  -p 3000:8080 \
  --add-host=host.docker.internal:host-gateway \
  -v open-webui:/app/backend/data \
  --name open-webui \
  --restart always \
  ghcr.io/open-webui/open-webui:main

# Open WebUI is now at http://localhost:3000
# Create your admin account on first visit
# Go to Settings → Connections → set Ollama URL to http://host.docker.internal:11434

Step 5: Use as a Private AI Server

The Mac mini M4 Pro's low power draw (30–45W typical, 92W peak) makes it ideal as an always-on private AI server. Connect to it from other devices on your network using Tailscale for secure remote access. The Ollama API on port 11434 and the MLX server on port 8080 are both OpenAI-compatible, so any tool that supports OpenAI can point to your Mac mini instead.

# Install Tailscale for secure remote access
brew install tailscale
sudo tailscaled &
tailscale up

# Your Mac mini gets a stable Tailscale IP (e.g., 100.x.x.x)
# From any device on your Tailscale network:
curl http://100.x.x.x:11434/api/generate -d '{
  "model": "qwen3.5:9b",
  "prompt": "Hello from remote device"
}'

# Set Ollama to listen on all interfaces (not just localhost)
launchctl setenv OLLAMA_HOST 0.0.0.0
# Then restart Ollama

Recommended Models for Mac mini M4 Pro 24GB

These models are tested and recommended for the 24 GB configuration. All are available via Ollama or as pre-quantized MLX models on Hugging Face.

Mac mini M4 Pro vs Alternatives

Understanding where the Mac mini M4 Pro fits relative to other options at similar price points helps confirm whether it is the right buy for your use case.

Power and Noise

The Mac mini M4 Pro draws approximately 30–45W under typical AI inference load and peaks at around 92W under sustained maximum load. It is completely silent under light load and nearly inaudible under heavy inference — the fan only becomes audible during sustained 100% GPU utilization. Running 8 hours per day at average 45W costs approximately $1.97/month in electricity at $0.15/kWh. This makes it one of the most efficient local AI machines available: you get 32B model capability at the power draw of a light bulb.

When to Upgrade to Mac Studio

The Mac mini M4 Pro is the right machine until you hit one of these ceilings: you need to run 70B models at Q8 quality (requires 64+ GB), you need faster inference on 32B models (Mac Studio M4 Max has 410 GB/s vs 273 GB/s), or you need to serve multiple concurrent users (Mac Studio's higher bandwidth handles parallel requests better). The Mac Studio M4 Max starts at $1,999 with 36 GB unified memory and 410 GB/s bandwidth — a $700 premium over the Mac mini M4 Pro 24 GB that is justified once you are regularly running 70B models or serving a small team. Buy Mac Studio M4 Max on Amazon (~$1,999+).