🧠 AI on your phone, with no internet
🌐 Français · English · العربية
Since 2025, you can run a real AI model (an "LLM") directly on an Android phone, offline, without sending your data to anyone. Few people know it. Here's how, and what you need to know to make it work well.
Why local AI?
| Cloud AI (ChatGPT, Gemini…) | Local AI on the phone | |
|---|---|---|
| Privacy | Your messages go to a server | ✅ Nothing leaves the phone |
| Internet | Required | ✅ Works in airplane mode, on a building site, on a plane |
| Cost | Subscription for the best models | ✅ Free |
| Power | Very high | More limited, but enough for many tasks |
Perfect for: summarizing a confidential document, rewording an email, translating offline, studying with a child without ads, or thinking out loud without it being stored anywhere.
1. Is your phone powerful enough?
The simple rule: a "4-bit" model takes about 0.6 to 0.7 GB of memory per billion parameters, plus some headroom.
| Phone RAM | Recommended model size | Example model families |
|---|---|---|
| 4 GB | 0.5 to 1 billion parameters ("1B") | Small Gemma, Qwen, Llama |
| 6 to 8 GB | 1 to 4 billion ("2B", "3B", "4B") | Gemma 3n, Qwen, Llama 3.2 3B, Phi |
| 12 GB and more | up to 7 or 8 billion | "7B / 8B" models in 4-bit |
You'll find your RAM in Settings → About phone. Start small: a model that's too big makes the phone very slow or crashes the app.
2. The easy way: a ready-made app
- Google AI Edge Gallery: Google's experimental app to try models (including Gemma) offline: chat, questions about an image, transcription. Open source on GitHub (
google-ai-edge/gallery). - PocketPal AI: an open-source app that downloads GGUF models from Hugging Face and runs them locally.
- MLC Chat: another open-source demo app (the MLC LLM project).
In these apps: pick a small model, download it over Wi-Fi (from a few hundred MB to several GB), then switch to airplane mode to check it really works offline.
3. The geek way: llama.cpp in Termux
For developers who want full control (see the Termux guide):
pkg install git cmake clang
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build && cmake --build build --config Release -j 4
# download a .gguf model (e.g. a Q4_K_M version), then:
./build/bin/llama-cli -m ~/model.gguf -p "Explain the SSH protocol in 3 sentences."llama-server even starts a small local web interface you can open in the phone's browser.
4. Picking the right model file
- Q4_K_M: the best size/quality trade-off for a phone.
- Q8: more accurate, but about twice as heavy.
- An "instruct" or "it" model is made for chatting; a "base" model is not.
- Prefer models that are good in your language: not all of them speak French or Arabic as well as English.
5. Limits to keep in mind
- A small model makes more mistakes and confidently invents facts. Always check what matters (figures, laws, medication, structural calculations).
- It doesn't know the news: its knowledge stops at its training date.
- It heats up and uses battery: plug the phone in for long sessions.
🧪 AI in my apps
Several of my apps use AI (image enhancement, background removal, data processing), sometimes directly on the phone. Join the club to test them first, with free Premium.