An AI assistant that works with no internet (laptop and phone)
A small ChatGPT-style AI that lives on your own laptop and phone, so you can still ask it to explain, draft, summarise or translate when the internet is down.
01When it saves you
If the internet is out for days, ChatGPT and every other online AI stop working. A local AI can still help you write a letter to your landlord or insurer, turn a long instruction leaflet into simple steps, work out quantities, or explain something you've found in offline Wikipedia. Nothing you type leaves the device, so it's also private.
02How it works
An AI model is one big file, the AI's "brain". A free app downloads the model once, then runs it on your own processor. After that it needs no internet at all.
Small models (a few GB) are what normal laptops and phones can run. Google's Gemma 4 E2B is a good current one: about 4.6GB in Ollama on a laptop, or about 3.1GB as a phone-sized file. It's good at writing, rewording, summarising and simple explanations. It is much weaker than ChatGPT at facts, maths and anything specialist, and it can be confidently wrong.
On a laptop, Ollama is the simplest route: one installer, and on Windows and Mac it has its own chat window. On a phone, PocketPal AI (free, open source, iPhone and Android) downloads a model and chats with it offline.
03What to buy
Both are free, need no account and work offline once the model is downloaded. Ollama is what our lesson uses; PocketPal AI is on both the App Store and Google Play.
Where: Ollama from [ollama.com/download](https://ollama.com/download). PocketPal AI from the store links on its [official GitHub page](https://github.com/a-ghorbani/pocketpal-ai).
Jan is a free, open-source chat app for Windows, Mac and Linux that downloads and runs models itself. Google AI Edge Gallery runs Google's small models offline on Android 12+ and iOS 17+, but it is a beta showcase app.
Where: Jan from [jan.ai](https://jan.ai). AI Edge Gallery from the links on its [GitHub page](https://github.com/google-ai-edge/gallery).
Avoid
- "Offline AI" apps with a subscription that are really just a front end for an online service. If it doesn't download a model of several GB, it isn't running on your phone.
- Models marked cloud in Ollama's list (for example
gemma4:cloud). They run on Ollama's servers and stop working with no internet. - A model too big for your memory. On a laptop with 8 to 16GB of RAM, stay with
gemma4:e2b; bigger ones crawl or freeze the machine. - GPT4All, which older guides recommend: its last release was February 2025.
04You’ll also need
- Laptop: 8GB of RAM minimum, 16GB recommended, and about 7GB of free disk space (the Windows installer alone is about 1.6GB). Check RAM in Windows under Settings, System, About, Installed RAM.
- Phone: a recent phone with plenty of memory. Gemma 4 E2B at phone size needs about 4GB of memory while it runs, on top of what the phone itself uses, so an older phone with 4GB of RAM won't cope. You also need about 3GB of free storage.
- Wi-Fi for the downloads, done now.
- Charging: running a model works the processor hard, so keep a power bank handy.
→Set it up
The short version is below. The full lesson has every screen.
- Laptop: install Ollama
Download it from ollama.com/download ↗ and run the installer. On Windows and Mac it opens a chat window.
- Laptop: get the model
Choose gemma4:e2b in the app's model list, or open a terminal and type
ollama pull gemma4:e2b(4.6GB). The lesson linked below walks through this step by step, with real speeds: about 26 to 29 tokens (word pieces) a second on a normal laptop with no graphics card. - Phone: install PocketPal AI
Install it from the App Store or Google Play using the links on its official GitHub page. Open Menu, Models, pick a model that fits your phone and tap Download, or tap + to add one from Hugging Face. Gemma 4 E2B at Q4_K_M is about 3.1GB.
- Load it and ask something
On the phone tap Load, then ask a short question. On the laptop, type in the chat window.
✓Test it before you need it
Turn Wi-Fi off on the laptop and Airplane mode on the phone. Ask each one: "Rewrite this as three simple steps: turn off the water at the stopcock under the sink, open the cold tap to drain it, then call a plumber." If you get an answer on both, they work offline.
⚡On the day
- Use it for words: letters, lists, plans, summaries, rewording instructions, simple translations.
- Never use it for medicine doses, first aid or anything safety-critical. Check those in the NHS Medicines guide and WikiMed in Kiwix, and call 111 or 999 when you can.
- On a phone, load the model only when you need it and close it after. It uses a lot of memory and battery while it's running.
↻Keep it ready
!Limits, honestly
- Small models make things up. Treat every fact, number and name as a guess until you've checked it in a real reference such as offline Wikipedia.
- It knows nothing after its training date and can't look anything up: no news, no weather, no live information.
- Speed depends on your machine. A normal laptop is fine for short answers; an old laptop or phone can be slow.
- Bigger, cleverer models need far more memory: Gemma 4 E4B needs about 5.5 to 6GB and the 12B model about 7 to 8GB at the usual size, which most phones can't spare.