How on-device AI summarization works
Short version: modern browsers can run a language model locally. Zengist reads your files in the browser, sends the text to that local model, and only uses a cloud model when your browser has none. Every result tells you which happened.
What happens when you press the arrow
Five steps, in order. Steps one to three never touch a server.
- Your files become text on your device. PDFs are read with pdf.js, Word files with mammoth, images and scanned pages with on-device text recognition. Nothing is uploaded to do this.
- The text is chunked to fit the model. Small on-device models read a few thousand words at a time. Long material is split on paragraph boundaries, each part is reduced to dense notes, and the notes are combined into the final brief.
- An engine is chosen. Chrome's built-in model if it is ready, otherwise a private model you have already downloaded, otherwise the cloud relay.
- The brief is written. The same instructions are used on every engine: a TL;DR, key points, every date, amount and action found, and one line per source.
- The result stays with you. It is shown, kept in your browser's local storage for the recent list, and never sent anywhere.
The three engines
Chrome built-in AI (Gemini Nano)
Since Chrome 138, desktop Chrome and Edge ship a compact model called Gemini Nano behind the Summarizer API, and since Chrome 148 the more general Prompt API is available to websites too. The browser downloads the model once, about 2 GB, and shares it across every site that asks for it. Requirements are a desktop operating system (Windows 10 or later, macOS 13 or later, Linux, or a Chromebook Plus), 22 GB of free disk space, and either a GPU with more than 4 GB of memory or 16 GB of RAM with four cores. Phones are not eligible yet.
This is the best on-device option. Google tunes the summarizer specifically for this job, and the model runs fast on ordinary laptops. The context window is small, so Zengist chunks anything longer than a few pages.
A private model over WebGPU
WebGPU lets a web page use your graphics chip directly. As of 2026 it is on by default in Chrome, Edge, Safari 26 (Mac, iPhone and iPad) and Firefox on Windows and Apple Silicon. With it, Zengist can download an open model, Qwen 2.5 1.5B on computers (about 1 GB) or Qwen 2.5 0.5B on phones (about 350 MB), and run it entirely in your browser using Transformers.js. The download is cached, so the second run starts immediately. It is always opt-in: the size is shown and nothing downloads until you agree.
Quality is honest rather than magical. The 1.5B model writes a clean brief from a few pages; the 0.5B phone model is fine for short material and loses detail on long, dense text. When you need the strongest result on a big pile, the cloud path is better, and you can choose it in settings.
The cloud fallback
Older browsers, most Android phones, and machines without the disk space for Chrome's model use a cloud model instead. Your text is sent to the Zengist server, relayed to the model provider, streamed back, and dropped. There is no database and no history on our side. The path is capped per connection and per day so it can stay free, and the app says so plainly if a cap is reached.
How it compares
Most free summarizers differ on three things: whether your file leaves your device, whether you need an account, and how many files you can combine into one brief.
| Tool | Files leave your device? | Account | Many files at once | Price |
|---|---|---|---|---|
| Zengist | No when on-device (default); yes on the cloud fallback, not stored | None | Up to 20 sources | Free |
| ChatGPT, Claude, Gemini (chat apps) | Yes, always | Required for files | Yes | Free tier with limits |
| Typical "free PDF summarizer" sites | Yes, always | Often after one or two uses | Usually one | Free with upsell |
| Desktop apps with local models (LM Studio, Ollama) | No | None | Depends | Free, needs setup |
The desktop apps are the closest cousins. They give you full control and bigger models, at the cost of installing software and choosing weights yourself. Zengist is the zero-setup version of the same idea, tuned for one task.
What "nothing is stored" actually means
On-device runs never make a network request with your content. The only requests are the page itself and, if you opted in, the model download from a public CDN, which sends nothing of yours to anyone. On cloud runs, the server holds the text in memory for the length of the request and writes nothing to disk. The model provider's own retention terms apply for that request; they are named in the privacy policy rather than glossed over.
Limits worth knowing
- Up to 20 sources per run, 50 MB per file, 10 links per run.
- Scanned PDFs are read up to 30 pages; the text layer of normal PDFs has no page limit.
- The cloud path accepts about 60,000 words per run. Beyond that, split the material or run on-device, where chunking is unlimited.
- Password-protected PDFs and encrypted archives are not read.
Tips for better briefs
- Write the task. "What do I owe and by when" beats a generic summary of a bill.
- Group related files in one run. The per-source line at the bottom shows what each contributed.
- Pick Long in settings for contracts and reports; Short for emails and articles.
- On a laptop, turn on Chrome's built-in AI once and every later run is private and quick.
Try it or read why it exists.