Outlier  ›  learn

How to use AI offline on your Mac

Quick answer
  • Most "AI" apps are a thin client for a server — switch off Wi-Fi and they die. That's not offline AI.
  • Genuinely offline AI = the model weights are files on your disk, and inference runs on your own Apple Silicon chip. Nothing is fetched per query.
  • The setup: download one signed Mac app, pick a model that downloads to your disk, then turn Wi-Fi off and keep working.
  • Verify it yourself: enable Airplane Mode and watch it still answer. Works on flights, in dead zones, in rooms where nothing's allowed online.

Here's the test that settles it. Open whatever AI app you use, turn on Airplane Mode, and ask it anything. If it spins and then apologizes, it was never AI on your Mac. It was a webpage pointed at a data center, and you were the thin client the whole time. Real offline AI passes the airplane test: the model is sitting in a folder on your disk, the math runs on your own chip, and it answers with the radio off. This is how you set that up.

Why most AI needs the internet

The big assistants don't run on your computer. ChatGPT, and most chat apps wrapped around it, are clients: you type, the text travels to a server farm, a model you never touch produces a reply, and it travels back. The intelligence lives in someone else's building. That's why a dropped connection breaks it, why there are usage caps, and why "is the model down?" is a question that exists at all. Your device is doing roughly as much thinking as a TV remote.

Local AI flips the arrangement. The model (the actual file of weights) gets downloaded to your Mac once, and after that the work happens on your Apple Silicon chip. No request leaves the machine. The internet was only ever the delivery truck; once the model is parked in your garage, you don't need the truck again. Open-weight models (the kind published openly, like the ones on HuggingFace) make this possible because the weights are yours to keep, not a service you ping.

The offline setup (four steps)

No terminal, no Docker, no account. On an Apple Silicon Mac (M1 or newer, macOS 12+):

  1. Download the signed app, once. Grab Outlier: one notarized .dmg from the site, drag it to Applications. This is the only thing that needs the internet, and only this once.
  2. Pick a model and let it land on your disk. Open the app and choose a tier. Nano and Lite are free and small; they download to a folder on your Mac and stay there. Bigger tiers exist if you want them, but you don't need the cloud to use what you've downloaded.
  3. Ask one question while still online, just to confirm the model loaded and you're getting answers. Nothing's offline yet; you're checking the engine turns over.
  4. Cut the cord. Turn Wi-Fi off (or flip on Airplane Mode), then keep chatting. The answers come from the model already on your disk, computed on your chip. No reconnect, no "you're offline" wall.

That's the whole thing. The reason it works is unglamorous: a model is a file, your chip can do the arithmetic, and a file doesn't need a signal.

How to verify it's really offline

Don't take anyone's word for it, including ours. There's a hardware-level proof you can run in ten seconds, and it's the same test that exposes the fakes.

Run the same test on a cloud app and it fails immediately, which is the point. Offline isn't a setting you trust; it's a thing you can see. Outlier runs with Wi-Fi off because the model is genuinely on your machine, not as a feature toggle, but because there's nowhere else for the work to happen.

Receipts: Inference (the per-query work an assistant does when you talk to it) is roughly 80–90% of AI's compute load, per MIT Technology Review (2025). Cloud AI ships that work to a data center every single message; offline AI does it on the chip you already own, which is why the network can be off. Outlier's weights are publicly inspectable on HuggingFace: open files, not a remote service.

What changes when you're offline (and what doesn't)

Being honest about the trade keeps this trustworthy. Offline, two things change. Anything that needs live data (today's news, a current stock price, a fresh web search) isn't there, because there's no web to reach; the model answers from what it learned, not from the internet. And speed depends on your chip: a heavy local model like Core 27B runs about 20.7 tok/s on an M1 Ultra where cloud flagships run 80–100, so big models feel deliberate rather than instant.

What doesn't change is everything that matters for getting work done. Writing, editing, coding, summarizing a document you already have, reasoning through a problem, drafting an email: all of it runs the same with the radio off. And the privacy isn't a promise, it's physics. With Wi-Fi off, your prompts can't go anywhere because there's no server to send them to. No usage cap, no queue, no "you've hit your limit." The same model is sitting there on a flight at 36,000 feet, in a windowless room where nothing's allowed online, or in the dead zone where your bars vanish. The marginal cost of an offline answer is your wall socket: no data center required.

Frequently asked questions

Can ChatGPT work offline?

No. ChatGPT is a thin client for OpenAI's servers — the model never lives on your device, so every message is a round-trip to a data center. Turn off Wi-Fi and it stops working entirely. Offline AI means the opposite: the model weights are files on your own disk and inference runs on your Mac's chip, with no network call per query.

Does offline AI send my data anywhere?

No. When the model runs on your machine and Wi-Fi is off, your prompts and answers physically cannot leave the device — there is no server to send them to. Your conversations stay on your disk, in your backups, under your control. That's the strongest privacy guarantee there is: not a policy promise, a wiring fact. (More on that in why local AI keeps your code private.)

Is offline AI slower than ChatGPT?

It depends on the model and your chip. Local Core 27B runs around 20.7 tok/s on an M1 Ultra where cloud flagships run 80–100, so heavy models feel slower than a fast data center. But there's no network latency, no queue, and no usage cap — and on a flight or in a dead zone, a slower answer beats no answer.

What do I need to run AI offline on a Mac?

An Apple Silicon Mac (M1 or newer) on macOS 12 or later, and enough free disk space for the model you pick — sizes range from about 2.4 GB up. You download the signed app and the model once over the internet; after that, no connection is required. No account, no terminal, no Docker.

Try Outlier free

Free Nano + Lite — local, private, no account. Pro $20/mo or $149/yr adds everything (all 7 model tiers incl. Plus 397B). Lifetime Pro from $99 (Founding 200, first 200 seats) or $200 (Founders 500). Apple Silicon only.

Download for Mac