Ollama-darwin.zip installer from ollama.com or run brew install ollama via Homebrew in Terminal. Launch Ollama.app to initialize Apple Metal acceleration, then run ollama run llama3.2 to start chatting with your first local model.
Apple Silicon (M1, M2, M3, and M4 chips) has established MacBooks and Mac Studios as the premier developer hardware for running local AI models. Thanks to Apple’s Unified Memory Architecture (UMA) and high-bandwidth Metal GPU acceleration, a Mac Studio or MacBook Pro with 64GB–128GB of RAM can run massive 70B parameter models that would otherwise require multiple enterprise NVIDIA GPUs.
This setup guide walks through both installation methods for macOS: Standalone GUI App (with Menu Bar daemon) and Homebrew CLI Service (ideal for developers using terminal dotfiles and background services). For a complete overview of CLI commands and architecture, check out our Ollama Complete Guide.
1. System & Hardware Compatibility Matrix
| Apple Silicon Chip Tier | Unified Memory (RAM) | Max Recommended Model Size | Expected Speed (Q4 Quantization) |
|---|---|---|---|
| Base M1 / M2 / M3 / M4 | 8 GB – 16 GB | 3B to 8B (Llama 3.2, Qwen 2.5 7B) | ~35 – 55 tokens/sec |
| M-Series Pro (M2/M3/M4 Pro) | 18 GB – 36 GB | 14B (Qwen 2.5 14B, DeepSeek-R1 14B) | ~45 – 70 tokens/sec |
| M-Series Max (M2/M3/M4 Max) | 36 GB – 128 GB | 32B to 70B (Qwen 32B, Llama 3.3 70B) | ~30 – 45 tokens/sec (70B) |
| M-Series Ultra (M2/M3 Ultra) | 64 GB – 192 GB | 70B+ & DeepSeek-R1 70B | ~50 – 65 tokens/sec (70B) |
2. Method 1: Standalone Application Installation (Recommended)
Step 1: Download and Install Ollama.app
- Download
Ollama-darwin.zipfrom the official Ollama website. - Double-click the zip archive to extract Ollama.app.
- Drag
Ollama.appinto your macOS Applications folder. - Launch Ollama from Launchpad or Spotlight (Cmd + Space).
Step 2: Install CLI Symlinks
When launched for the first time, Ollama will prompt: “Install command line tool?”. Click Install and enter your macOS administrator password. This creates the symlink at /usr/local/bin/ollama.
Step 3: Run a Model in Terminal
Open Terminal or iTerm2 and execute:
ollama run llama3.2Ollama will stream the model layers, automatically bind to Apple Metal GPU compute kernels, and open an interactive chat session.
3. Method 2: Homebrew Installation (Developer / Headless Mode)
If you manage your developer tools using Homebrew, install Ollama without the GUI wrapper:
brew install ollama
brew services start ollamaThis registers Ollama as a macOS launchd background service that automatically boots on login.
4. Apple Silicon Memory Optimization & Environment Variables
By default on macOS, models are stored in your user home directory at:
~/.ollama/modelsTo persist environment variables (like OLLAMA_MODELS for external Thunderbolt SSDs or OLLAMA_HOST=0.0.0.0 for LAN access) with the GUI app on macOS, set them via launchctl:
launchctl setenv OLLAMA_MODELS "/Volumes/FastSSD/OllamaModels"
launchctl setenv OLLAMA_HOST "0.0.0.0:11434"
launchctl setenv OLLAMA_ORIGINS "*"For full instructions on configuring storage paths, see our Ollama Model Storage Guide and the complete Ollama Environment Variables Guide.
5. macOS Diagnostics & Metal Verification
To verify Metal GPU utilization on macOS, open Terminal while a model is generating and run:
ollama psYou can also open Activity Monitor (Cmd + Space → Activity Monitor), navigate to the Window → GPU History menu, and watch the Metal GPU activity graph spike to 90–100% during token generation.
If you experience memory pressure warnings or slow generation, read our Ollama GPU Configuration Guide and our Ollama Troubleshooting Guide.
Frequently Asked Questions
Can I run Ollama on older Intel Macs?
Yes. Ollama supports Intel-based Macs with AVX2 CPU instructions. However, inference will run entirely on the CPU or discrete AMD GPU, which is noticeably slower than Apple Silicon Metal unified memory.
How do I completely quit Ollama on macOS?
Click the llama icon in the top macOS Menu Bar and select Quit Ollama. If running via Homebrew, run brew services stop ollama.
How do I delete downloaded models on macOS?
List installed models with ollama list, then delete any model with ollama rm <model-name>. See our step-by-step tutorial on How to Delete and Remove Ollama Models.