Apple Silicon Macs are the easiest devices for running local LLMs: unified memory lets the GPU access all RAM, so a 16GB MacBook runs 7B models smoothly and a 64GB Mac can run 34B models that would need a high-end PC. This guide covers install, model selection, and tuning.

💡 What You Will Learn

Apple Silicon Macs are the easiest devices for running local LLMs: unified memory lets the GPU access all RAM, so a 16GB MacBook runs 7B models smoothly and a 64GB Mac can run 34B models that would ne

📜 Table of Contents

• Why Mac
• Install
• Pull and run a model
• Model selection
• Performance and parameters
• Integrations
• FAQ

Run Ollama Locally on Mac: Complete Guide for Apple Silicon

Apple Silicon Macs are the easiest devices for local LLMs: unified memory means CPU and GPU share all RAM, so a 16GB MacBook runs 7B models smoothly and a 64GB Mac can run 34B-class models. Order: install, pull models, pick the right size, tune.

Why Mac

Traditional PCs depend on VRAM: 8GB VRAM handles 7B, 16GB barely handles 13B. Apple Silicon has no separate VRAM: 8GB RAM runs 1-3B, 16GB runs 7-8B, 32GB runs 13-14B, 64GB runs 30-70B quantized. Actual limits depend on quantization and context length.

Install

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

Check with ollama --version. The local service runs on port 11434 by default.

Pull and run a model

ollama pull llama3.1:8b
ollama run llama3.1:8b

Other commands: ollama list, ollama ps, ollama rm <model>, ollama serve. See the official Ollama library for available models.

Model selection

1. Match size to RAM: 16GB -> 7-8B, don't force 14B.
2. Prefer quantized builds (Q4/Q5): smaller, faster, minimal quality loss.
3. When unsure, start with a 7-8B instruct model (Llama 3.1 8B, Qwen2.5 7B).

Performance and parameters

Reference: 16GB MacBook ~30-50 tok/s on 7B; 64GB Mac ~5-12 tok/s on 70B quantized. Key env vars: OLLAMA_CONTEXT_LENGTH (bigger = more RAM), OLLAMA_NUM_PARALLEL, OLLAMA_KEEP_ALIVE. GPU acceleration uses Metal automatically.

Integrations

Ollama exposes an OpenAI-compatible API at localhost:11434, so VS Code extensions (Continue, Cline), chat frontends (Open WebUI) and curl scripts all work out of the box.

FAQ

Q: Battery drain? A: ~15-20% per hour for 7B chat on a MacBook Air; idle models are released automatically. Q: Not enough RAM? A: Smaller model, shorter context, or heavier quantization. Q: Vs cloud API? A: Free, offline, private; weaker than flagship cloud models for hard reasoning. Use both.