You can summarize text on your phone without internet in three ways: Apple Intelligence’s Writing Tools on an iPhone 15 Pro or newer, Gemini Nano’s summarization on supported Android phones, or a local AI app that runs a language model on the device. The built-in options are quick for a paragraph or an email. A local AI app gives you more control, handles your own prompts and documents, and works on older phones too.
All three keep the text on your phone for most requests, which is the whole point when you’re summarizing a lease, a medical letter or work notes. Those are exactly the files you shouldn’t be uploading to a cloud chatbot.
Option 1: Apple Intelligence on iPhone #
On an iPhone 15 Pro or later with Apple Intelligence turned on, select text in almost any app, open Writing Tools, and choose Summary, Key Points, List or Table. Apple’s roughly 3-billion-parameter on-device model handles many of these requests without a connection. Larger requests can go to Apple’s Private Cloud Compute servers, which need internet.
Good for: emails, articles in Safari, notes, quick summaries in any app. Limits: you can’t write your own instructions, it needs a recent iPhone, and longer text may need the cloud.
Option 2: Gemini Nano on Android #
On supported Android phones, including the Pixel 9, 10 and 11 series and the Galaxy S25 and S26 series, Google’s on-device Gemini Nano model can summarize text in apps that use it. Google’s developer documentation lists summarizing articles and conversations as bulleted lists among its on-device features, and says it works without a network connection.
Good for: summaries inside apps that support it, such as recording and messaging apps on supported phones. Limits: only on recent flagships, and only in apps that have added the feature.
Option 3: A local AI app #
A local AI app downloads an open language model once and runs it on your phone. You paste text or attach a file and ask for exactly the summary you want.
Personal LLM works this way on iOS and Android. Paste text into a chat, or attach a PDF, text or Markdown file; the file is split into passages and searched on the phone, and the model answers from the matching ones while telling you which it used. Nothing is uploaded, there’s no account, and it all works in airplane mode once a model is downloaded. A 4B model is a good balance for documents on most phones; if yours has 6 GB of RAM or more, Qwen 3.5 9B gives the most careful answers in the catalogue, and Ministral 3 3B suits long documents on smaller phones. See how much RAM you need to work out which fits.
Good for: custom summaries, long documents, older phones, anything private. Limits: smaller models than the cloud, and very long documents need the approach below.
Which option should you use? #
| Apple Intelligence | Gemini Nano | Local AI app | |
|---|---|---|---|
| Phones | iPhone 15 Pro and later | Recent Android flagships | Most phones with 3 GB+ RAM |
| Works offline | For many requests | Yes | Yes |
| Custom instructions | No | Depends on the app | Yes |
| Attach PDFs | Via share and select | Depends on the app | Yes |
| Choose the model | No | No | Yes |
Prompts that make better summaries #
The real advantage of a local AI app is that you decide what “summary” means:
- Quick version: “Summarize this in three sentences.”
- Skimmable: “Give me five bullet points, most important first.”
- Action items: “List every task, who owns it and any deadline.”
- Decisions: “What was decided, and what’s still open?”
- Plain language: “Explain this to someone with no background in the topic.”
- Numbers only: “List every figure, date and amount mentioned.”
- For a reply: “Summarize this email, then suggest a short response.”
- Pros and cons: “Summarize the arguments for and against.”
Say who the summary is for and how long it should be. “A two-line summary I can text to my partner” gets a very different answer from “a detailed brief.”
How do you summarize a long PDF on a phone? #
This is where offline summarizing differs most from the cloud, so it’s worth understanding the mechanics.
A model can only take in so much text at once, its context window. Some models support very long contexts on paper; Ministral 3 3B in the Personal LLM catalogue has a 256k window. But on a phone a big context eats memory and takes a long time to process, so practical limits are far below the specs. A 40-page contract won’t fit in full.
Local document chat therefore works like a search engine plus a reader. The app splits the document into chunks of roughly 300 tokens, runs a keyword search (BM25) to find the chunks matching your question, and gives only those to the model. That’s fast and light on memory, with one consequence: the model answers from the passages it was shown, not from the whole file.
So the question you ask decides how well it works.
- Ask something specific first. “What is the notice period to end this lease?” lands on the right passages. “Summarize this” on a 40-page file only ever sees a sample.
- Name the section. “Summarize the section on termination and penalties” points the search where you want it.
- Work section by section for a full summary. Ask for the headings, summarize each in turn, then ask the model to combine those summaries.
- Start a new chat for each long document so earlier text doesn’t eat into the context.
- Check the source. The app tells you which passages it used, so you can read the original lines before relying on the answer.
For short documents, a two-page letter or a one-page policy change, a single “summarize this in five bullet points” is fine. How to chat with a PDF offline goes further into file types and what the search misses.
If what you want is a confusing official letter put into everyday words rather than condensed, that’s a different job with different prompts. See explain a document in plain English.
What about scanned PDFs? #
A scanned PDF is a set of images with no text layer, so there’s nothing for a document search to read. If a PDF was printed and then scanned, or photographed, you have two choices:
- Use a vision model. Photograph or screenshot each page and ask a vision-capable model to read and summarize it. This works well for one or two pages.
- Run text recognition first. iPhone’s Live Text and Android’s built-in text recognition copy text out of an image. Paste that text into the chat.
Can you trust an AI summary? #
Mostly, with checks. For plain, well-structured documents a 4B to 9B model summarizes accurately most of the time. Mistakes cluster in three places:
- Numbers. A figure copied wrong, or attached to the wrong item.
- Negations. A missing “not” or “unless” flips the meaning.
- Tables. Structure often collapses when text is pulled out of a PDF.
Before acting on a summary, check key numbers, dates and names against the original, and ask “What did you leave out?” to see whether something important was cut. For contracts, medical or legal documents, treat the summary as a way to find the right paragraph faster, not a substitute for reading it or asking a professional. Why AI makes things up explains why this happens and how to reduce it.
Asking the model to quote the lines it used is the single best habit. It makes invented answers easy to spot.
Summarizing recordings, transcripts and news #
AI summarizers work on text, so for a meeting or lecture you need a transcript first. Transcribe the recording on your phone with an on-device speech-to-text app, then paste the transcript in and ask for a summary, action items or a list of questions raised. Our guide to offline speech to text covers the options.
Local models also don’t know recent events, so they can only summarize news you give them. That limit, and how to work around it, is in does offline AI know current events.
Frequently asked questions #
Can I summarize text without internet? #
Yes. Apple Intelligence Writing Tools on iPhone 15 Pro and later, Gemini Nano on supported Android phones, and local AI apps like Personal LLM all summarize on the device. Local apps work on a wider range of phones.
Can AI summarize a PDF without internet? #
Yes. An app that runs the language model on your phone reads and summarizes a PDF with no connection, once the model is downloaded. Cloud chatbots and most PDF readers with AI features need a connection because the model runs on their servers.
Why does my offline AI miss parts of a long PDF? #
Phone models hold only a limited amount of text at once, so local apps search the document and show the model the passages matching your question. Ask specific questions, or summarize one section at a time, to cover the whole file.
Is it safe to upload a contract to ChatGPT? #
It’s stored on OpenAI’s servers with your conversation and handled under its privacy policy, which can include use for training unless you’ve turned that off. For documents with personal, financial or confidential details, an offline tool avoids sending them anywhere.
Is AI summarization accurate? #
Usually good enough for the gist, but it can miss or blur details. Check numbers, dates and anything you’ll act on against the original text, especially with small on-device models.
Can a phone AI read a scanned document? #
Not through text search, because a scan has no text layer. A vision-capable model can read a clear photo of a page, or you can use your phone’s built-in text recognition to copy the text out first.