On 13 October, Wharton professor Ethan Mollick pointed out that 19 AI models small enough to run on a modern smartphone now beat GPT-4, the model that led the field for a whole year.
You download a model over wifi once, so it answers in airplane mode, using your phone’s own chip.
That makes it the right tool for the jobs you would rather no company saw:
- Private notes. A health worry, a money mess or a family row, sorted into something you can act on.
- A sensitive message. The text to your landlord, your ex or your boss, calmed down before you send it.
- No signal. On a plane, abroad without data, or anywhere the bars drop to nothing.
Pick an app
All three are free and run the model on the phone itself. Store pages read 13 October 2026.
Google AI Edge Gallery
- Works on: iPhone (iOS 17 or later) and Android, updated in October.
- Models: Google’s Gemma 4 family, plus other open models you can import.
- Worth knowing: the answers are worked out on the phone, but the iPhone app’s privacy label says it collects diagnostics and a device ID, and its Agent Skills, such as Wikipedia look-ups, need the internet.
- Pick it if: you want Google’s own app on either phone.
PocketPal AI
- Works on: iPhone (iOS 15.1 or later) and Android, updated in October. It is open source, built by an independent developer, and its code is public on GitHub, with a new release on 4 October.
- Models: open models from Hugging Face, the big public library of AI models, including Gemma 4.
- Worth knowing: its App Store label says no data is collected. Its optional web search and remote-server features do send text off the phone, so leave them off for private work.
- Pick it if: your iPhone is a few years old, or you want the widest choice of models.
Locally AI by LM Studio
- Works on: iPhone and iPad only (iOS 18.1 or later), updated 14 September.
- Models: Llama 3.2, Gemma, Qwen, Apple’s own on-device model and more, with Gemma 4 only on phones with enough memory.
- Worth knowing: its App Store label says no data is collected. Its LM Link feature runs bigger models on your computer instead, which needs a connection.
- Pick it if: you have a recent iPhone and only want an iPhone app.
Set it up and test it in airplane mode
- Free up space first. A model is a single big file: Google’s model page lists the smaller Gemma 4 phone model, E2B, at about 2.6 GB, and a download can fail partway on a full phone.
- Download on wifi and start with the smallest model the app suggests. A model too big for your phone freezes or crashes; both PocketPal and Edge Gallery have a built-in speed test (a benchmark) that shows what your phone can handle. If the app asks you to sign in to Hugging Face before a download, that account is free.
- Switch on airplane mode, then ask a test question. If it answers, the model is running on your phone, and this is the one check you can do yourself.
- Leave off anything that goes online. That means web search, remote servers, LM Link and online Agent Skills, and skip any feedback option, which shares what you send with the app maker. For anything really private, keep airplane mode on while you type.
- Start a new chat for each job. Phone models hold far less text at once than ChatGPT or Claude, so an old conversation crowds out your new note.
The prompt: sort a note or rewrite a message, privately
Small models follow short, labelled instructions far better than a long, loose brief, so this prompt gives the model one job per run and a tight set-up to fill in.
Pick SORT for a messy note, REWRITE for a message you are nervous about sending, or CHECK to see how a message will land before you change it. Fill in the set-up, paste your text at the bottom, and the model checks its own answer before it shows you.
On an older phone, delete the two jobs you are not using before you paste, so the model only reads the one it needs.
You are my private writing helper, running on my phone with no internet. I use you for notes and messages I would not paste into an online chat, so one wrong name, date or number could cause real trouble. Work only from the text I give you, and never add a fact. SET-UP (I fill this in) JOB: [pick one: SORT, REWRITE or CHECK] WHO IT IS FOR: [e.g. "just me", "my landlord", "my sister after an argument"] TONE: [e.g. "calm and firm", "warm but short", "plain"] LENGTH: [e.g. "under 120 words", or "any"] MUST KEEP: [anything that must stay word for word, e.g. "the date 4 November and the flat number", or "nothing"] MY TEXT: [paste the note or message here] THE JOBS SORT: Turn my messy note into three short lists. FACTS: what happened or what I know. TO DO: each task with who and when; if either is missing, write "not stated". QUESTIONS: anything open or unclear. Keep my own words where you can. REWRITE: Rewrite my message for the person above, in the tone I asked for. Keep every fact, date and number exactly as I wrote it. Cut repetition and blame. If the reader would need something my text does not say, write [NEED: what is missing] instead of guessing. CHECK: Leave the message as it is. Read my message as the person receiving it would. List up to three lines that could land badly, say why in one short line each, and offer a calmer version of each. HOW TO WORK 1. Read all of MY TEXT once before you write anything. 2. If MY TEXT is empty or JOB is not filled in, ask me one short question and stop. 3. If a line in my text could mean two things, quote it and ask me which I meant. 4. Do the one job I picked, and only that job. 5. Never invent a name, date, number, place or promise. Use "not stated" or [NEED: ...] instead. 6. Short sentences, plain words, no extra headings. BEFORE YOU ANSWER, CHECK - Every name, date and number in your answer appears in my text. - The MUST KEEP items are exactly as I wrote them. - You stayed within my LENGTH. Then end with one line: "Least sure about: ..." naming the part I should check myself.
Reading on a computer? Email the prompt to yourself, or open this page on your phone and tap Download .txt.
What a good answer looks like
- Your note: “dentist rang moved to thurs?? ask mum re lift. pay deposit before fri. jen owes me for tickets”
- Facts: The dentist appointment has moved, possibly to Thursday.
- To do: Ask Mum about a lift (who: me; when: not stated). Pay the deposit (who: me; by Friday). Get the ticket money back from Jen (who: me; when: not stated).
- Questions: Is the dentist definitely Thursday? How much does Jen owe? (amount not stated)
- Least sure about: whether “thurs??” means the new day is confirmed.
The honest bit
- It makes more mistakes. A phone model is far smaller than the ones behind ChatGPT or Claude, so check every name and number, starting with the line the prompt says it is least sure about.
- It knows nothing new. There is no web, so it cannot look up today’s news, a price or opening hours, and anything it says about recent events is a guess.
- It is slower and hungrier. Answers take longer on an older phone, the chip works hard enough to drain your battery faster, and each model takes gigabytes of storage.
- It can drift. If it repeats itself, does every job at once or changes a date, start a new chat, delete the unused jobs and paste a shorter text; if it still struggles, try the next model up, as long as the app’s benchmark says your phone can take it.
- Check the first answer. Compare it with your own note before you rely on it.
Try it on tonight’s to-do list
Install one app on wifi tonight, download the smallest model it offers and save the prompt in your phone’s notes app, so it is two taps away when you have no signal. Then switch on airplane mode and run SORT on a note you already have.
Send this page to the friend who never pastes anything personal into ChatGPT.
A few quick questions
Will it work on my old phone?
Maybe, with a small model. PocketPal’s Android page asks for 6 GB of memory (RAM) for smaller models and 8 GB or more for better ones. On iPhone, PocketPal runs on iOS 15.1 or later, the oldest iOS version of the three apps. Start with the smallest model and run the app’s benchmark.
Do these apps cost anything?
All three are free to download and the models are free. PocketPal sells optional add-on personas it calls Pals, which you do not need for the prompt above.