- Learn
- Get started
Ollama on a Mac: setup for Apple silicon, step by step
Install Ollama on an M-series Mac, pick a model that fits your memory, check it's using the GPU, and change the settings that matter.
A Mac with Apple silicon (M1 or newer) is the easiest computer to run local AI on. There are no drivers to install. Ollama uses the GPU through Metal automatically, and the GPU shares the Mac's unified memory, so a 32 GB Mac can load models a 16 GB graphics card can't.
If you're choosing which Mac to buy, read Macs for local LLMs. This guide is about setting up the one you have.
Step 1: Install Ollama
Download Ollama for macOS from ollama.com, open it, and drag it to Applications. Start it once. It adds an icon to the menu bar and runs a small server in the background, and it offers to install the ollama command for Terminal.
Step 2: Pick a model for your memory
macOS lets the GPU use about two thirds of memory on Macs up to 36 GB, and about three quarters above that. Leave room for your other apps too.
| Your Mac's memory | Start with | Download |
|---|---|---|
| 8 GB | qwen3:4b | 2.5 GB |
| 16 GB | qwen3:8b | 5.2 GB |
| 24 to 32 GB | gpt-oss:20b | 14 GB |
| 36 to 48 GB | qwen3:30b | 19 GB |
| 64 GB and up | llama3.3:70b | 43 GB |
Step 3: Run it
Open Terminal and run the model for your memory, for example:
ollama run qwen3:8bType a question at the >>> prompt. Type /bye to quit. The Ollama cheat sheet has the rest of the commands.
Step 4: Check it's on the GPU
ollama ps100% GPU is what you want. A split like 20%/80% CPU/GPU means the model is too big for the share of memory macOS gives the GPU. Drop to a smaller one.
You can also open Activity Monitor and look at the GPU History window while it answers.
How fast it will be
On a Mac, the chip matters more than the memory size for speed. A base M4 with 16 GB, from our calculator:
And an M4 Pro with 48 GB, which has more than twice the bandwidth:
On the M4 Pro, gpt-oss 20B and Qwen3 30B-A3B are estimated at about 48 tok/s, while the dense Qwen3 32B is about 8. Mixture-of-experts models read only a small part of their weights per token, which suits Macs well. See mixture of experts explained.
Settings worth changing
Ollama on a Mac reads its settings from environment variables set with launchctl. For example, to move models to an external drive:
launchctl setenv OLLAMA_MODELS /Volumes/External/ollama-modelsThen quit Ollama from the menu bar and open it again. The same pattern works for OLLAMA_HOST and OLLAMA_KEEP_ALIVE. Settings made with launchctl setenv are lost on restart, so you'll need to run it again after a reboot. See where Ollama stores models.
Giving the GPU more memory. On recent macOS versions you can raise the share of memory the GPU may use with sudo sysctl iogpu.wired_limit_mb=<megabytes>. It resets on reboot. Leave several GB for macOS itself, or the whole machine slows down.
Other good Mac apps
- LM Studio has a model browser and chat window, and can run models in Apple's MLX format, which is often a little faster on Macs than GGUF. See Ollama vs llama.cpp vs LM Studio.
- Open WebUI gives Ollama a ChatGPT-style window. See set up Open WebUI.
Intel Macs
Older Intel Macs run Ollama on the processor only. Stick to qwen3:4b, and read running AI without a GPU for what to expect.
Where to go next
- New to this? Try your first week with local AI.
- Words like quant and context confusing? See the local AI glossary.
- Compare Macs: the M4 Max and other chips each have their own page.
One email when there's a new guide, podcast episode, or result from the rack. No spam.