Gemma 4 on iPhone: Free AI, Offline, No Subscription

TL;DR

• Google has a free app, Google AI Edge Gallery, that runs its Gemma 4 model directly on your iPhone.

• No account, no subscription, no credit card. The app is free and so is the model.

• Once the model is downloaded, it works in airplane mode. Your data never leaves the phone.

• This is not ChatGPT. It's smaller, slower, and does less. That's normal, and it isn't the point.

• Bonus: the model doesn't even know it's running on your phone. That's the most useful lesson in the whole exercise.

---

Most people think using AI means sending your question to a data centre somewhere in Virginia, which answers two seconds later. That's true for ChatGPT. It's true for Gemini. It isn't mandatory.

Since February 2026, Google has published an iPhone app called Google AI Edge Gallery. It runs a real AI model on your phone's own processor. No server, no account, no bill. And since June, it officially supports Gemma 4, Google's most recent open model.

First: Gemma is not Gemini

This is the number one confusion, and it's a fair one. Both names start with G and come from the same company.

Gemini is Google's commercial product. It runs on Google's servers, it's built into Gmail and Search, and the bigger versions cost a subscription.

Gemma is Google's family of open models. “Open” means Google publishes the whole model: anyone can download it and run it on their own machine, for free. Gemma 4 shipped on April 2, 2026 under the Apache 2.0 licence, a permissive licence that even allows commercial use.

Both come out of the same research. Gemini is the car with a driver. Gemma is the engine, handed to you to put wherever you want.

What you need

• An iPhone running iOS 17 or newer.

• Realistically an iPhone 13 or newer. It installs on older ones, but it will crawl.

• Free space: 2.59 GB for Gemma 4 E2B, 3.66 GB for the beefier E4B.

• Wi-Fi, once, for the initial download.

How to install it, in five minutes

The App Store listing: developer Google, free, no in-app purchases. Always check the developer name.

Five models on offer. The app flags Gemma-4-E2B-it “Best overall” itself. Note “supports text input”: on iPhone, it's text only.

1. Open the App Store and search “Google AI Edge Gallery”. Check that the developer says Google.

2. Install it. Free, no in-app purchases.

3. Open the app and go to AI Chat. Five models are offered.

4. Pick Gemma-4-E2B-it: the app itself flags it “Best overall”, and it's the lighter one at 2.59 GB. On a newer iPhone, E4B at 3.66 GB answers better.

5. Download over Wi-Fi.

6. When it's done, put your phone in airplane mode and ask it a question.

Actually do that last step. It's the moment it clicks.

The best test: ask it where it is

I tried it. Airplane mode, no connection at all, and I asked the model whether it was “offline”.

Its answer: it isn't offline the way an unplugged computer is, it said, because it “runs on computer servers”.

Top of the screen: “Model on GPU”. Below it, the model insists it runs on computer servers. Both cannot be true. The phone is the one telling the truth.

That's false. It was running on my iPhone's GPU with no internet access whatsoever. The app says so in plain text two lines above the answer: “Model on GPU”. The response took 11.6 seconds, generated in my hand.

The model doesn't know. It has no way to know. It was trained on billions of texts where AI assistants write “I run on servers”, so that's what it generates. It isn't lying. It simply has no self-knowledge to consult.

A language model generates plausible text. It does not look things up in a database of facts. Including about itself.

If you take one thing from this article, take that. It explains half the errors you'll see using AI, and it's exactly the kind of misunderstanding that makes AI projects fail in the first 90 days. It cost you nothing to learn.

Let's be honest about what it does, and doesn't

What works well: summarizing text, rewriting an email, answering general questions, kicking ideas around.

What works less well, and it needs saying:

• On iPhone, these variants take text only. No images, even though Gemma 4 is multimodal elsewhere. The context window is 32K, not the 256K of the big versions.

• Answers are slow. Budget roughly ten seconds for a paragraph.

• On older iPhones the app can freeze after a few exchanges. Users report it, including on an iPhone 14.

• Agent Skills, which lets the model take actions on the phone, is still very limited.

• The model is small. On complex reasoning or long documents, it falls behind.

None of that is a product flaw. It's physics. You're running a model on a phone battery instead of a server rack. The miracle is that it works at all.

Why you should care, even if you're not a geek

First, it's the cheapest training that exists. The best way to learn AI is reps on real work, and the cost of starting is the only real barrier. Here that cost is zero. No credit card, no subscription to cancel, no 14-day trial that turns into a bill.

Second, your data stays with you. Nothing leaves the phone. For a Quebec SMB handling client information, this is the clearest possible demonstration of what “local processing” means. Law 25 requires you to know where the personal information you hold ends up. Here the answer is one word: nowhere.

Third, it demystifies. Once you've watched an AI model get its own nature wrong, in your hand, in airplane mode, the magic drops away. What's left is a tool. That's exactly where we want you, and it quietly settles several myths that cost small businesses money.

The ten-minute test

Take an email you have to write anyway. Airplane mode. Ask the model to draft it. Look at what's good, what's wrong, and what you'd have said differently.

You've just learned more about AI than any hour-long webinar would teach you.

FAQ

Is it really free?

Yes. The app is free on the App Store with no in-app purchases, and the Gemma models are published under the Apache 2.0 licence. There's no account to create. The only cost is storage space on your phone.

Does Google see my questions?

No. Once the model is downloaded, all processing happens on the phone. The only time the app touches the internet is to download the model in the first place. After that, airplane mode and it works the same. That's the simplest way to verify it yourself, and ironically the model itself will tell you otherwise.

Does this replace ChatGPT for my business?

No, and that isn't the goal. A 2-billion-parameter model on a phone is not in the same league as a frontier model in a data centre. Treat it as a free, private test bench, not a production tool.

Where to go from here

If you want to put numbers on what automation is worth in your business before talking to anybody, run your own figures through the ROI calculator. And if you'd rather just listen, Simon breaks down one AI story a week, in plain language, on Cogeco Outaouais 104.7 FM.

Automate. Delegate. Eliminate. But before any of that: try it.