How to run an AI assistant entirely on your iPhone

Two years ago this meant a hobby project and a warm phone. Today it is an app download. Here is what the phone needs, what to expect, and how to prove to yourself that nothing is leaving.

What you need

• An iPhone 15 Pro or newer. Running a language model locally takes 8 GB of memory. The iPhone 15 Pro and Pro Max have it, and so does every iPhone 16 and iPhone 17 model. Older phones do not, and no setting changes that.
• iOS 26.
• About 4 GB of free storage. The model is a one-time download of roughly 3.5 GB. If you want voice conversations, the neural voice adds about 330 MB, and the extended voice pack for more languages another 400 MB.
• Wi-Fi for the first download. After that, the model works with the radios off.

How it works, briefly

A language model is a very large set of numbers, called weights, and a program that runs them. Orvena downloads the weights for Qwen 3.5 4B once, then runs them with MLX, Apple's framework for machine learning on Apple silicon. The weights are compressed to four bits per number so the whole model fits in memory alongside iOS. The result is an assistant that reads your message, reasons, and writes its answer on the phone's own processor, the same hardware that renders a game.

The same model reads images. Attach a photo and it describes what it sees without uploading the picture anywhere.

What to expect

Speed. On an iPhone 16 Pro, Orvena writes at roughly fifteen words a second, which is faster than most people read. Harder questions can be asked to think first, which adds a few seconds before the answer starts.

Quality. A four-billion-parameter model is very good at the things people ask an assistant to do: schedule, rephrase, summarise, look something up, work through a plan, read a screenshot. It is weaker than a data-center model at long essays and at obscure facts it was never trained on. For current information, ask it to search. The query goes to the search provider; the reading and the answer happen on the phone.

Battery and heat. A short question costs almost nothing. A long voice conversation warms the phone the way a game does and uses battery accordingly. Nothing runs in the background unless you schedule it.

Step by step

1. Install Orvena from the App Store.
2. Open it. The model download starts with a progress indicator. Stay on Wi-Fi.
3. Ask something. Permissions for calendar, reminders, photos, or health are requested only when a request needs them.
4. Turn on airplane mode and ask again. This step is the whole point of the article.
5. Optional: unlock voice, and the neural voice downloads too.

The airplane-mode test

It is the only test that matters, because it does not depend on reading a policy. With the radios off, an app can only answer with what is on the phone. Ask a question, attach a photo, move a meeting. If all of that works, you have verified where the model runs in a way no privacy page could show you.

When you might still want a cloud model

Some work needs a bigger model: a long document, a specialised domain, a demanding piece of writing. Orvena lets you add one with your own API key from OpenRouter, and it asks for consent per provider, and again before sensitive history would leave the phone. The point is that this becomes a decision you make for a specific request, not the default condition of using an assistant.