AI automation
Use AI offline on your phone (local LLM)
In short
Yes, an AI model can run entirely on your smartphone – no internet, account or subscription: with a free app like PocketPal AI or MLC Chat you download a small model (e.g. Gemma or Llama, 1B class) once and then chat even in airplane mode. Your inputs never leave the device – the trade-off is smaller models and slower answers.
With the well-known assistants like ChatGPT, the actual AI model runs in the cloud – your phone only sends the request back and forth. A local LLM is the opposite: the whole model sits on your device and does the computing right there. The big advantage is maximum privacy – neither your inputs nor the answers leave the phone, and it works on a plane, in a tunnel or with the network fully off. The trade-off: these are smaller models. They're good for summarising, drafting, quick questions and as an offline reference – but they don't replace a large cloud model for tricky reasoning. For beginners it's an exciting, free way into the topic of "AI that you own yourself".
Requirements
- Enough memory (RAM): This is the deciding factor – not the app. Small 1B models run on phones with 6–8 GB of RAM, larger 3–4B models want a flagship with 8–12 GB.
- A reasonably recent phone: Android or iPhone from roughly 2021 onwards is enough for light models. A high-end device isn't needed to get started.
- Free storage space: Depending on size, the model files are around 1–4 GB and are downloaded once.
- Realistic expectations on speed & battery: Answers come more slowly than from the cloud (roughly 15–20 seconds for a short reply), and longer sessions use noticeable battery.
Good to know: downloading the model needs internet once. After that everything runs offline – you can switch the phone to airplane mode and keep chatting.
Free
- Install a free offline AI app – for beginners, PocketPal AI (Android and iPhone) or MLC Chat work well.
- In the app, select and download a small model. A good starter model is a 1B model like Gemma 3 1B, Llama 3.2 1B or Qwen2.5 1.5B.
- Wait for the download (internet needed once). After that you can switch the phone to airplane mode.
- Simply ask a question in the chat window – the answer is generated entirely on the device.
- If the app crashes or gets very slow, the model is too large for your phone – then pick a smaller one.
Summarise the following text in five simple bullet points and explain difficult terms in a short aside: [PASTE TEXT]
Limits: Small models make more mistakes on complex reasoning and answer more slowly than the cloud. For factual questions, better double-check the answer.
For advanced users
- If your phone has enough RAM (8–12 GB), pick a larger 3–4B model – it gives noticeably better answers for only a little more memory use.
- Look for models in GGUF format on Hugging Face and load them into the app – that gives you the full choice instead of only the suggested models.
- Watch for 4-bit versions (Q4): they save memory; for more complex tasks, choose a higher quality level.
- If you like, use a terminal solution on Android (e.g. Ollama) to manage models in a desktop-like way.
Rewrite this email to be more polite and shorter without changing the content. Then give me one version in a casual tone and one in a formal tone: [PASTE TEXT]
Added value: Larger models give noticeably better answers and more context – with full privacy. Ideal for sensitive texts the cloud shouldn't see.
Frequently asked questions
What phone do I need for this?
No high-end device: an Android or iPhone from around 2021 with 6–8 GB of RAM is enough for 1B models. Memory is the deciding factor – larger 3–4B models want a flagship with 8–12 GB.
Is a local model as good as ChatGPT?
No – cloud models are clearly stronger at complex reasoning. For summarising, rewriting, quick questions and offline reference, the small models are surprisingly usable though. Many people use both: local for sensitive things, cloud for the rest.
Why is this more private than normal AI apps?
Because the entire model computes on the device: neither your inputs nor the answers are sent to a server. Only the one-time model download needs internet – after that everything works in airplane mode.