Yes, a Mac with Apple silicon can run an AI model with the Wi-Fi switched off, using an app called Ollama. You go online once to download the app and a model, and after that everything runs on the laptop. In our test an 8 GB Mac handled a small model well, and 16 GB is the memory size we would recommend.
The numbers below are ours, measured on one Apple M2 Mac with 8 GB. We did not test Intel Macs or Windows laptops.
What you need
Ollama's macOS page lists macOS Sonoma (v14) or newer. It supports Apple M series chips with CPU and GPU, or x86 processors with CPU only.
Disk space depends on the model. Ollama warns that models can be tens to hundreds of GB in size, and it keeps them in the ~/.ollama folder. The two small models we ran took 2.0 GB and 3.3 GB.
Set it up while you are online
Download Ollama from its website. The documented way to install is to open the downloaded file and drag the Ollama app into your Applications folder.
When you open it, the setup prompts offer a choice between cloud models and a local one. Pick the local one, because Ollama's cloud runs models in its own cloud, not on your Mac, so a cloud model stops working the moment your connection does.
Then open Terminal and download a model once:
ollama pull llama3.2
ollama run llama3.2
The second command opens a chat in the terminal. Type a question and press Return. Type /bye to leave. The ollama run llama3.2 line comes from the model's own page in Ollama's library.
Ollama's quickstart now points to a newer model, Gemma 4 E2B, with a download of about 7.2 GB. It recommends 8 GB of available VRAM, or unified memory on a Mac. We have not run that model. Our own reading is that a Mac with 8 GB in total will not have 8 GB free, because the system and a browser already take 4 to 5 GB. Ollama adds that with less memory it can use system RAM, but responses may be slower.
Which model size fits 8 GB and 16 GB
These are the models we ran, with the memory each used while answering.
| Model | On disk | Memory used | What we saw on the 8 GB Mac |
|---|---|---|---|
| Llama 3.2 3B | 2.0 GB | 2.5 GB | First words after about 2.5 seconds, 29 word pieces per second. Fast enough. |
| Gemma 3 4B | 3.3 GB | 3.5 to 3.7 GB | On a busy Mac, 33 seconds to the first word. On a quiet one, about 5 seconds for a full answer. |
| A 7B to 8B model | not tested | about 5.5 to 6.5 GB (estimate) | Not measured. Plan on a 16 GB Mac. |
Our conclusion from the test: a Mac with Apple silicon (M1 or newer) and 16 GB works. On 8 GB, the 3B model works only if few other apps are open, and a 4B model was already too slow.
The Gemma result shows why one run proves little. On the busy day the Mac was already using 7.5 GB of swap, which is memory borrowed from the disk, and Gemma crawled. With little else running, the same model answered in about 5 seconds. The 3B model stayed fast on both days, so close your browser before you judge a model.
We also tried models around 1B. They are not usable as an AI assistant. The 7B to 8B figure is an estimate, not a measurement.
Check that it really works offline
"It runs on my laptop" is easy to claim and easy to get wrong, because Ollama can also reach its cloud. Do this once, before you need it.
- While online, run the model once and ask it anything, so you know it works.
- Switch Wi-Fi off in the menu bar and unplug any network cable. Open a web page to confirm that it fails.
- Quit Ollama and open it again, so nothing is left over from the online session.
- Run
ollama run llama3.2and ask a question. It should answer. - In a second Terminal window, run
ollama ps. Ollama's FAQ says it shows which models are currently loaded into memory.
If step 4 fails with a connection error, you probably picked a cloud model, or the download did not finish. Ollama's FAQ documents a local only mode that disables its cloud features. The catch, in its own words, is that you lose Ollama's cloud models and web search. We have not tried the setting ourselves.
On privacy, Ollama says it does not see your prompts or data when you run locally.
The limits
A small model can be wrong in ways that sound sure. We asked a Dutch question about how many drops of bleach go into 4 litres of water. Llama said 2 drops, and the library it was reading said 8.
Gemma told us a closed fridge keeps food cold for 48 hours. The library said 4 hours, and 48 hours is a full freezer. That is why our food timer quotes the USDA, not a model.
Across 33 survival questions, Llama got 26 fully right and Gemma 22. Llama made 3 serious safety mistakes and Gemma 4. The language you ask in changes the results too. English was best, with 10 of 11 fully correct answers for both models. Dutch managed 8 of 11 for both, and French 8 of 11 for Llama and only 4 of 11 for Gemma. The full results are in our offline AI test.
A model on its own also cannot look anything up, so pair it with documents you can check. Our guide to an offline survival library covers what to keep. And a laptop that is answering questions drains its battery faster, which the laptop runtime calculator helps you plan for.
To remove a model you no longer need, run ollama rm followed by its name, for example ollama rm llama3.2.
We're building Waystone, an offline survival assistant that is still in development, on this same kind of setup: a small local model and a library of official pages. If you try it yourself first, treat every number the model gives you as a claim to check.