Your phone is already powerful enough to run a real language model with no internet connection. Small open models now answer questions, summarize text, help with writing and describe photos entirely on the device, and free apps make setup a few taps. This guide covers which apps to use on Android and iPhone, which models fit, what hardware you need, and the jobs that still need a computer.
| Quick answer: to run AI locally on your phone, install a free on-device AI app such as Google AI Edge Gallery (Android and iPhone) or PocketPal AI (Android and iPhone), download a small model such as a 1-4B Gemma, Qwen or Phi build while you are online, then switch to airplane mode and chat. A phone with 8GB of RAM or more gives the best results. Video generation is not practical on phones yet. |
Table of Contents
What a phone can and cannot do offline
On-device AI works well for text and simple vision tasks: chatting, answering questions from the model’s own knowledge, summarizing or rewriting text you paste in, drafting emails, translating short passages and describing what is in a photo. Everything is processed on the phone, so it works on a plane and nothing you type is sent to a server.
The limits come from memory and heat. Phones run small models, usually 1-4 billion parameters, which are less capable than cloud chatbots on hard reasoning and have no live web access. Long chats slow down, and sustained use warms the phone and drains the battery. Generating AI video is far beyond what phones can do locally today; for that you need a desktop graphics card, as our guide on how to generate AI videos locally explains. For desktop options, see our ranking of the best local AI video generator models.
The best apps to run AI locally on your phone
How we compare: we include apps that run models on the device rather than calling a cloud service, judged on platforms, model choice, cost and ease of setup. Details come from each project’s own app listing or repository.
| App | Platforms | Models | Cost | Best for |
|---|---|---|---|---|
| Google AI Edge Gallery | Android, iPhone | Google’s Gemma family and other compatible models | Free, open source | The easiest start, with chat and image tasks |
| PocketPal AI | Android, iPhone | Any small GGUF model from Hugging Face | Free, open source (MIT) | Choosing your own models |
| MLC Chat | Android | Models compiled for the MLC LLM engine | Free, open source | Tinkerers who want GPU-accelerated inference |
| Private LLM | iPhone, iPad, Mac | Curated list of open models | Paid one-time | A polished iPhone app with no setup |
Google AI Edge Gallery
Google’s AI Edge Gallery started as an Android experiment and is now on Google Play and the App Store. It downloads models such as Gemma to the phone and runs them offline for chat, questions about images and other tasks. Reviews report that the largest models can make the app freeze on phones with less memory, so start with a smaller one.
PocketPal AI
PocketPal AI is a free, MIT-licensed app built on llama.cpp, the same engine behind many desktop tools. It lets you browse and download small GGUF models from Hugging Face inside the app, so you can try Qwen, Phi, Gemma or Llama builds and keep the one that runs best on your phone. Gated models need a free Hugging Face access token.
Built-in options from Apple and Google
Recent phones also ship with on-device models. Apple Intelligence runs a model of roughly 3 billion parameters on the iPhone 15 Pro and newer, and Apple lets developers use it offline through its Foundation Models framework. On Android, Google’s Gemini Nano runs on-device on supported phones. These power built-in features such as writing tools and summaries rather than a free-form chat app, but they show how much phones can now do without the cloud.
What you need on your phone
| Phone memory (RAM) | Model size that runs well | What to expect |
|---|---|---|
| 6GB | 1-2B parameters | Short answers and simple tasks; larger models may crash |
| 8GB | 2-4B parameters | Useful everyday assistant for writing and questions |
| 12GB or more | 4-8B parameters | Noticeably better answers, with slower replies on bigger models |
Models are stored on the phone, so keep a few gigabytes of free space for each one. Pick 4-bit quantized builds, which are much smaller with modest quality loss. A recent flagship or upper mid-range chip makes replies much faster. If the app offers GPU or NPU acceleration, turn it on. Our explainer on what a small language model is covers why these compact models work so well.
Which model to download first
Start with the smallest model the app recommends for your phone, test it on the tasks you actually need, then try one size up. Small Gemma models are the default in Google’s app and handle chat and image questions well. Small Qwen and Phi builds are popular in PocketPal for writing and reasoning. Keep two models at most: a fast small one for quick questions and a slightly larger one for better answers. Our best local LLM guide compares these families in more depth.
Step by step: your first offline chat
On Android: install AI Edge Gallery or PocketPal AI from Google Play, open the models list, download a small model over Wi-Fi, then start a chat. On iPhone: install either app from the App Store and do the same. In PocketPal, search Hugging Face from the Models screen and pick a 4-bit build that the app says fits your phone.
Then turn on airplane mode and send a prompt. If the reply arrives, the model is running locally. Keep prompts short at first, close other heavy apps, and if replies are slow or the app crashes, switch to a smaller model. Delete models you no longer use to recover storage.
Privacy, battery and when to use a computer instead
Local models keep your text on the device, but check each app’s settings for optional analytics or cloud features, and download apps only from the official stores or repositories. Expect noticeable battery use during long sessions and some warmth on the back of the phone.
Move to a computer when you need larger models, long documents, coding help or media generation. A laptop running Ollama handles models several times bigger than a phone can, and our local AI guide explains the whole setup, from first install to video generation.
How one nurse used offline AI on night shifts
Illustrative example. Leila Haddad is a nurse in Lyon who wanted help rewording handover notes and checking drug-name spellings, but hospital policy ruled out pasting anything into cloud chatbots, and the ward’s signal is poor.
She installed PocketPal AI on her 8GB Android phone, downloaded a small 4-bit model at home over Wi-Fi and confirmed it worked in airplane mode. She uses it for general wording and never enters patient identifiers, keeps a second, smaller model for faster replies, and checks anything clinical against official sources. The app needs no signal and her text never leaves the phone.
Frequently Asked Questions
Can I run AI locally on my phone without internet?
Yes. Apps such as Google AI Edge Gallery and PocketPal AI download a model to your phone once, then run it entirely offline. You only need internet to install the app and download models; after that, airplane mode works.
What is the best app to run AI locally on Android?
Google AI Edge Gallery is the easiest start, with Gemma models and image tasks built in. PocketPal AI is the best choice if you want to pick your own models from Hugging Face. Both are free and open source.
Can I run AI locally on an iPhone?
Yes. Google AI Edge Gallery and PocketPal AI run models offline on iPhone, and Private LLM is a paid option. Apple Intelligence also runs an on-device model on the iPhone 15 Pro and newer for built-in features.
How much RAM does a phone need to run a local LLM?
About 6GB runs 1-2B parameter models, 8GB is comfortable for 2-4B models and 12GB or more handles 4-8B models. Use 4-bit quantized builds and close other apps for the smoothest results.
Can I run a local AI video generator on Android?
Not in a practical way today. Open video models need a desktop graphics card with roughly 8-16GB of dedicated video memory and plenty of system RAM. Phones can run small text and image models; for video, use a PC with a capable GPU.
Is local AI on a phone private?
The model runs on the device, so prompts and replies are not sent to a server. Check the app for optional analytics or cloud features, install it only from official stores or repositories, and remember that anything saved on the phone is only as secure as the phone itself.
Conclusion
To run AI locally on your phone, install AI Edge Gallery or PocketPal AI, download a small 4-bit model over Wi-Fi and test it in airplane mode. A phone with 8GB of RAM or more gives the best experience for writing help, questions and image descriptions.
Use a computer for bigger models and anything involving video. Start small, and keep the model that answers fast enough on your own phone.
