In short: a modern smartphone can run its own neural network that works in airplane mode and sends nothing to any server. For iPhone the easiest app is Locally AI from the makers of LM Studio; for Android and iPhone — PocketPal. Both are free. A model weighs from one and a half to three and a half gigabytes and is downloaded once; after that you don’t need the internet. Below: how to install it, which model to pick for your phone and what to expect from a pocket neural network.
What a neural network on a phone can do
A pocket model is smaller than a desktop one, but it is enough for everyday things:
- rewrite a message to be politer or shorter, suggest a reply;
- translate a text or a menu while travelling — without roaming or Wi-Fi;
- explain an unfamiliar word, summarise an article, sketch a plan;
- describe what is in a photo — for models marked Vision.
What not to expect: fresh news (the model knows the world only as of its training date), error-free facts and long complex reasoning. Also, under load the phone gets noticeably warm and uses battery — that is normal.
iPhone: Locally AI by LM Studio
Locally AI by LM Studio is a free app from Element Labs, the developer of LM Studio. It runs on iPhone and iPad with iOS 18.1 or newer and on Mac with macOS 26, and runs models on Apple’s MLX engine. There is no Android version.
- Install the app from the App Store and open it.
- Choose your first model. On first launch the app offers a list. On iPhone 15 Pro, all iPhone 16 models and newer there is an Apple Foundation option — the same model that powers Apple Intelligence: you don’t need to download it, it is already in the system.
- Look into the model catalogue if you want something smarter. The app labels the sizes and suggests which iPhones each model is recommended for. For example, in the Bonsai family from PrismML the app, at the time of writing, offers the 4B variant (about 2 GB) for iPhone 15 Pro and newer, and the 9B one (about 3.5 GB) for iPhone 17 Pro, iPhone Air and newer.
- Keep the app open while the model downloads — the app warns about this itself.
- Type. After the download you can switch on airplane mode: answers will keep coming.
The labels in the catalogue: Thinking — the model “reasons” before answering (more accurate but slower), Vision — it understands images, High CPU Usage — heavy for this phone, it will warm it up and drain the battery.
Bonsai models have ternary weights: each takes one of three values (−1, 0 or +1). That makes them noticeably more compact than ordinary models of the same size and a good fit for phones in particular.
Android and iPhone: PocketPal
PocketPal AI is a free open-source app, available in both Google Play and the App Store. No sign-up needed. Its main difference is that it can download any model in GGUF format straight from Hugging Face, the largest store of open neural networks, not only from a built-in list.
PocketPal: a model from Hugging Face in five steps
- Install PocketPal from Google Play or the App Store.
- Open the models section and tap ”+” in the bottom-right corner.
- Choose “Add from Hugging Face” — a search across the store opens.
- Find a model by name and size, for example
qwen3.5 4b, pick a file and download it. - Load the model and start a chat. From then on everything works offline.
If a model on Hugging Face is gated by a licence agreement, PocketPal asks for an access token: you get one in your account settings on the Hugging Face website and paste it into the app’s settings.
For Android there is also an alternative from Google — AI Edge Gallery: an open app for Android 12 and newer, with Gemma models.
Which model to pick for your phone
A simple rule of thumb: the model file should fit into the phone’s RAM with plenty to spare — the system and open apps still need theirs. How much memory your phone has is shown in the settings (“About phone” on Android) or in your iPhone model’s specs.
| Model | Approximate file size | Who it suits |
|---|---|---|
| 2B (Bonsai 2B, Qwen 3.5 2B) | 1–3 GB | Older and budget phones: fast, but basic |
| 4B (Bonsai 4B, Qwen 3.5 4B) | 2–4 GB | Most phones of recent years — the best balance |
| 9B (Bonsai 9B, Qwen 3.5 9B) | 3.5–7.5 GB | Flagships with plenty of memory; an ordinary 6–7 GB 9B file is heavy for a phone, the ternary Bonsai is about half that |
Start with 4B. If it answers quickly and the phone doesn’t overheat, try 9B; if it lags, go back to 2B.
Models without refusals on a phone
Ordinary models refuse some topics. Through PocketPal you can install models without refusals — open neural networks with that behaviour cut out: in the Hugging Face search type 4B abliterated or 9B abliterated, for example Qwen3.5-9B-abliterated. Take GGUF versions — with GGUF in the file name. The essentials in short: fewer refusals, but no fewer mistakes, you still need to check facts, and such a model is not for children.
Tainet VPN hides both the contents and the metadata: from the network's side your traffic is indistinguishable from ordinary HTTPS, and the point owner can't see where you go. Sign up through the Telegram bot or in your account on this site, from $0.10 a day.
Try it →Phone or computer
A phone is always at hand and offline. A computer means models several times larger and smarter: a laptop with 16–32 GB of memory runs 9–27B models that come noticeably closer to cloud chatbots. If you need power rather than mobility, see how to run a neural network on your own computer.
So why the internet at all
A local model needs the internet once — to download the app and the model file, which is gigabytes, so Wi-Fi is better. After that your chat with the model stays on the phone. But everything else on the phone still goes online: cloud chatbots, messengers, the browser. How to protect that traffic on someone else’s Wi-Fi or while travelling — in the guide to VPN on a smartphone.
Frequently asked questions
Is it free?
Yes: Locally AI, PocketPal and the models in this article are all free. PocketPal has an optional one-time purchase with a voice mode; the app works fully without it.
Is Locally AI available on Android?
No, Locally AI is released only for iPhone, iPad and Mac. On Android use PocketPal or Google AI Edge Gallery.
Does the neural network work in airplane mode?
Yes. Once the model is downloaded, the app doesn’t use the internet — switch on airplane mode and check.
Why does the phone get warm and drain the battery?
A neural network loads the processor and graphics fully, like a demanding game. That’s normal. If it gets too hot, take a smaller model or one without the Thinking label.
What if the model is slow or the app crashes?
Most likely the model doesn’t have enough memory. Close background apps and try a model one size down: 4B instead of 9B or 2B instead of 4B.
A local neural network keeps everything you ask it to itself. But the traffic of your other apps — cloud chatbots, messengers, the browser — is visible to the hotspot owner on public Wi-Fi. Tainet hides both the contents and the metadata: sign up in the Telegram bot or in your account on this site, the first 7 days are free, then from $0.10 a day — one subscription for your phone and computer.
