Run Ollama on macOS in 2026: Full Setup Guide for Apple Silicon

Step-by-step macOS setup for Ollama, including Metal GPU acceleration and model picks.

## Run Ollama on macOS in 2026 Apple Silicon Macs are among the best local LLM machines: unified memory gives you big model capacity at high speed. Here is the full setup. ### Step 1: Install Ollama [bash] curl -fsSL https://ollama.com/install.sh | sh # or download Ollama.app from ollama.com for the menu-bar version [/bash] ### Step 2: Verify Metal acceleration [bash] ollama run llama3.2:3b --verbose # Look for Metal in the device list - this means GPU acceleration is active [/bash] ### Step 3: Pick models by RAM - **8 GB Mac**: qwen3:4b, llama3.2:3b - fast and useful - **16 GB Mac**: qwen3:7b, mistral:7b - the sweet spot, 15-30 tokens/s - **32 GB+ Mac**: qwen3:14b, llama3.1:8b, even 70B Q4 with offloading (slow) - **64 GB+ Mac**: 70B Q4 runs at 10-20 tokens/s - a local GPT-4-class setup ### Real speed expectations (community benchmarks, 2026) - M1/M2 16 GB: 7B Q4 about 15-20 t/s - M2 Pro/Max: 7B Q4 about 25-35 t/s, 13B about 15-20 t/s - M4 Max 128 GB: 70B Q4 about 15-25 t/s ### Tips - Close other apps during heavy inference - unified memory is shared - Use Ollama.app to auto-start and get the menu-bar controller - Set OLLAMA_KEEP_ALIVE to keep models hot for interactive use ### FAQ **Does Ollama use the GPU on Mac?** Yes, Metal acceleration is automatic on Apple Silicon. **What about Intel Macs?** They work but are 3-5x slower - Apple Silicon is strongly recommended for LLMs.