Run Ollama Locally on Mac: Complete Setup Guide for M1-M5 (2026)
From M1 Air to M5 Ultra: install Ollama on your Mac, pick models by RAM, integrate with editors, and set performance expectations.
💡 What You Will Learn
From M1 Air to M5 Ultra: install Ollama on your Mac, pick models by RAM, integrate with editors, and set performance expectations.
📜 Table of Contents
Run Ollama Locally on Mac: Complete Setup Guide for M1-M5 (2026)
Running LLMs locally means free, private, offline-capable AI. Thanks to unified memory, even a base M1 MacBook Air can run 7B models smoothly. Here's the full setup path.
Why Mac?
Unified memory acts as VRAM, so model size is limited by RAM, not GPU VRAM. Ollama uses Metal (GPU) acceleration by default and runs efficiently even on battery.
Install
- Official dmg from ollama.com (drag into Applications), or
brew install ollama
Verify with ollama --version.
Pick Models by RAM
| RAM | Model Tier | Use |
|---|---|---|
| 8GB | 1.5B-4B (quantized) | Chat, summaries, translation |
| 16GB | 7B-8B | General assistant, code completion |
| 24GB | 8B-14B (quantized) | Smarter assistant, long chats |
| 32GB+ | 14B-32B (quantized) | Complex reasoning, code gen |
| 64GB+ | 32B-70B (quantized) | Near-cloud experience |
Check exact sizes on the Ollama library; prefer quantized tags (e.g. q4) when RAM is tight.
Run It
ollama pull qwen2.5:7b
ollama run qwen2.5:7b
ollama list
/bye exits the chat. Swap the model name to switch models.
Three Ways to Use It
- CLI: interactive chat, great for developers.
- Web UI: run Open WebUI via Docker (
docker run -d -p 3000:8080 ... ghcr.io/open-webui/open-webui:main), open http://localhost:3000, and point the model service athttp://host.docker.internal:11434. - Editor: install the Continue.dev extension in VS Code and select Ollama as the provider—offline code completion and chat.
Performance & Tuning
- 7B models chat fluently on M-series (tens of tokens per second); measure on your own machine. 14B+ is noticeably slower.
- RAM tight? Switch to a smaller/more quantized model.
- Keep the service resident with
ollama serve(usually registered at login already). - Longer contexts use more RAM; keep context in check for long documents.
FAQ
Q: M1 Air 8GB? A: Yes—use 1.5B-4B models; 7B works but slowly.
Q: Why isn't the Neural Engine used? A: Ollama uses Metal (GPU), not ANE. Normal behavior.
Q: vs cloud models like ChatGPT? A: Small local models are weaker at complex reasoning, but fine for chat, summaries, translation, and code tasks. Think "free, private, good enough," not "flagship replacement."
Q: Where are models stored? A: ~/.ollama/models; remove unused ones with ollama rm <name>.
Note: Model lists and commands evolve; check the Ollama website and library.
❓ FAQ
M1 Air 8GB?
Yes—use 1.5B-4B models; 7B works but slowly.
Why isn't the Neural Engine used?
Ollama uses Metal (GPU), not ANE. Normal behavior.
vs cloud models like ChatGPT?
Small local models are weaker at complex reasoning, but fine for chat, summaries, translation, and code tasks. Think "free, private, good enough," not "flagship replacement."
Where are models stored?
`~/.ollama/models`; remove unused ones with `ollama rm <name>`. > Note: Model lists and commands evolve; check the Ollama website and library.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
