The two AI models behind our on-device tools, and when each one runs
Our AI tools, Summarize PDF and Ask PDF, are unusual in one specific way: the AI runs inside your browser, on your own device. Your document is never uploaded. Every other AI summarizer or chat-with-PDF tool sends your file to a server; ours has no server that receives a file at all.
Two different models make that possible. This post explains both in plain terms: what they are, when each one runs, what your device needs, where they are stored, and how many times they download. If you just want the short version: most people run the downloadable model, and you can pick which one you want on the tool itself.
The two models
1. The built-in model: Gemini Nano
Recent Chrome and Edge on desktop ship a small AI model inside the browser, called Gemini Nano. When it is present, our tools can use it directly through two browser features: the Summarizer API (for Summarize PDF) and the Prompt API (for Ask PDF).
- What it is: Gemini Nano, Google's on-device model, built into the browser.
- Parameters: a small model (a few billion parameters), tuned by the browser vendor. You do not choose or see the exact build; the browser manages it.
- What your device needs: a recent Chrome or Edge desktop with built-in AI enabled, plus the disk space and hardware the browser requires to provision it.
- Where it is stored: in your browser's own storage, managed by the browser, on your device.
- How it downloads: the browser downloads it once, in the background, and reuses it across every site and every session. We do not host it or control its size.
- Strengths: no download from us, and it is instant once your browser has provisioned it.
- Limits: it is a small model, so it is best at short, direct tasks. It is only present on some browsers, and only after the browser has downloaded it.
2. The downloadable model: Qwen2.5-1.5B
When the built-in model is not available (or you prefer it), our tools can run a model we host ourselves: Qwen2.5-1.5B-Instruct, through MLC web-llm. It runs on your GPU using WebGPU.
- What it is: Qwen2.5-1.5B-Instruct, an open-weights, instruction-tuned model from Alibaba's Qwen team, released under the Apache-2.0 license. It is multilingual and works well for summarizing and answering questions about text.
- Parameters: 1.5 billion parameters, 4-bit quantized (q4f16_1), with a 4,096-token context window.
- Download size: about 850 MB, one time.
- What your device needs: WebGPU (recent Chrome, Edge, Firefox, or Safari) and a GPU with roughly 1.6 GB of memory. The 4-bit build also uses the shader-f16 GPU feature; most modern GPUs have it.
- Where it is stored: in your browser's own storage for this site (IndexedDB), on your device. It never leaves your machine.
- How it downloads: only when you press Download. The weights come from this site, same-origin, never from a third party. After that it is cached, so it is a genuine one-time download and runs fully offline.
- How many times it downloads: once per browser. Summarize PDF and Ask PDF share the same model, so downloading it for one gives it to the other with no second download. It re-downloads only if you clear your browser's site data.
- Strengths: we host it, so we can tell you exactly what it is and whether it is already installed, and it is a bit larger than the built-in model, which helps on extraction-style questions.
- Limits: the one-time ~850 MB download, and it needs WebGPU.
When does each model run?
By default the tool picks automatically, in this order:
- Built-in Gemini Nano, if it is already downloaded and ready. This is instant and needs no download, so it wins when it is provisioned.
- Otherwise, the downloadable Qwen model, if your browser has WebGPU. Since WebGPU is common and a ready built-in model is comparatively rare, most people land here.
- Otherwise, the built-in model's own download, if there is no WebGPU.
- Otherwise, nothing. We never fall back to uploading your file.
This is a choice made once when the page loads, not a fallback after a failure. The tool does not try one model, see a weak answer, and switch; it commits to one.
Choosing the model yourself
You do not have to accept the automatic pick. On both tools there is a Model selector with three options:
- Auto: the order above.
- Built-in (Nano): force the browser's model, when your browser has it.
- Downloadable (Qwen): force our model, when your browser has WebGPU.
There is also a Check my system button that tells you, right there, whether your device supports the downloadable model (WebGPU and shader-f16), whether the built-in model is ready, and what GPU it found. If you pick a model your browser cannot run, the option is disabled and the check explains why.
A good rule of thumb: if the built-in model gives a weak or "cannot find it" answer on a document, switch to the downloadable Qwen model, which tends to do better on picking specific details out of text.
The one thing that never changes: no upload
Whichever model runs, the guarantee is identical. Your PDF's text is extracted in your browser and fed only to the local model. Nothing is uploaded, during the answer or after. The only network event either model can cause is a model coming down to you (the browser's model for Nano, or our weights for Qwen), never your document going up.
You can watch this yourself: every AI tool shows a live "0 B uploaded" indicator that measures outbound data in real time, and it stays at zero while you work. The model download is inbound only, so it never moves that number.
Where the Qwen model is hosted, and how it stays safe
The downloadable model is not pulled from some random third party at run time. We host it ourselves:
- Our own infrastructure. The weights (about 850 MB) live in our own AWS S3 bucket and are served over HTTPS through our CloudFront CDN, on the same origin as the rest of the site. At the moment you use the tool, nothing is fetched from an outside server.
- A known, open model. It is Qwen2.5-1.5B-Instruct, released by Alibaba's Qwen team under the Apache-2.0 license, in the standard MLC web-llm build. We serve those published weights as they are; we do not modify them.
- Sandboxed and boxed in. The model runs in a Web Worker, bound by the page's strict Content-Security-Policy. That policy's connect-src is limited to our own origin plus a couple of anonymous analytics endpoints that carry no file data, and there is no endpoint anywhere that accepts your document. So even the model's own code has no server it could send your file to.
- Checked on every change. Our code and infrastructure run automated security checks in CI before anything ships.
The built-in Gemini Nano model is downloaded and managed by your browser vendor under the same no-upload guarantee: your document is never part of any download, and it never leaves your device.
In short
- Two models, one promise: Gemini Nano (built into your browser) or Qwen2.5-1.5B (about 850 MB, downloaded from us and cached), both running on your device with no upload.
- Auto picks the fastest ready one; you can override it and check what your system supports on the tool.
- Qwen downloads once per browser and is shared between Summarize PDF and Ask PDF; Gemini Nano is downloaded and managed by your browser once.
Try them on Summarize PDF and Ask PDF.