How to Install Ollama on macOS (Step by Step)

Step 1: Install Ollama

Option A — The .zip (recommended for most people)

1. Go to ollama.com and click Download for macOS.
2. Open the downloaded zip and drag Ollama.app into Applications.
3. First launch: if Gatekeeper warns, right-click Ollama.app → Open (this is normal for open-source apps).

Option B — Homebrew (if you already use it)

brew install ollama

Then open Terminal and confirm:

ollama --version

Step 2: Run Your First Model

ollama run qwen3:8b

The first run downloads ~5GB and then drops you into a chat. Try something practical:

>>> Summarize this blog's RAG tutorial approach in 3 bullet points

Type /bye to exit. On an M-series Mac, the Metal GPU is used automatically — you'll typically see 20–60 tokens/sec on 8B models depending on your chip and RAM. Apple Silicon's unified memory is the big win: unlike a discrete GPU with fixed VRAM, the model can use whatever RAM you have, so a 64GB M4 Max runs quantized ~70B models comfortably.

Step 3: Model Management

Step 4: Use the Local API

Ollama runs an OpenAI-compatible server on http://localhost:11434:

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3:8b","messages":[{"role":"user","content":"Say hi"}]}'

Point any OpenAI-compatible tool at http://localhost:11434/v1 — the same pattern used in our 30-minute RAG tutorial and the privacy-first architecture of Lawyer Assistant. For Apple Silicon speed, some prefer MLX-based runtimes; we compare the options in Ollama vs llama.cpp.

Troubleshooting on macOS

• Gatekeeper warning — right-click → Open the first time; it's expected for zip-distributed apps.
• "ollama is not recognized" — reopen Terminal after install, or add ~/.local/bin to your PATH for the Homebrew install.
• Slow on Intel Mac — expected; Intel Macs run CPU-only. Consider a smaller model.
• Out of memory / swap thrash — the model is bigger than your unified memory. Use a lower quant or a smaller model; close memory-hungry apps (browsers!) before heavy runs.
• Want it to stop at login — quit from the menu-bar icon; it restarts on demand.

Frequently Asked Questions (FAQ)

How do I install Ollama on a Mac?

Two options: download the macOS .zip from ollama.com and drag Ollama.app to Applications, or run brew install ollama with Homebrew. Then open Terminal and run ollama run qwen3:8b — the first run downloads the model automatically.

Does Ollama use Apple Silicon (M1–M4) chips?

Yes — Ollama is highly optimized for Apple Silicon's unified memory, so the GPU is used automatically and models can use the full RAM pool (unlike discrete GPUs with fixed VRAM). An M-series Mac with 16GB unified memory runs 8B models comfortably and can stretch to ~30B models on 32–64GB machines.

Can I run Ollama on an Intel Mac?

Yes, but slower — Intel Macs run on CPU, so expect a few tokens/sec on 7–8B models. Apple Silicon is dramatically faster thanks to the Metal GPU acceleration.

Why does macOS warn about the downloaded app?

Gatekeeper warns about unsigned or newly downloaded apps. Right-click the app and choose Open (or go to System Settings > Privacy & Security and click Open Anyway) the first time. This is normal for open-source apps distributed as zips.

Where does Ollama store models on macOS?

Models live in ~/.ollama/models by default. Set the OLLAMA_MODELS environment variable to move them, and make sure you have ~2x the model file size free during download.

What's the best Ollama model for an M-series Mac?

For 8–16GB unified memory: qwen3:8b, phi4-mini, or gemma3:12b. For 32–64GB: qwen3:32b or gemma3:27b. Apple Silicon's unified memory means the "RAM = model size + context" rule applies — see our RAM guide for exact fits.

Sources

• Ollama — official site
• Ollama — GitHub (macOS install & FAQ)
• Ollama GPU documentation (Apple Silicon/Metal)
• Qwen3 — Ollama library
• Local AI Zone — GGUF model directory