Translate a PDF with on-device AIAI

No. The translation is produced by a model that runs inside your browser, so the document text is never sent anywhere. There is no endpoint that accepts your file. The live "0 B uploaded" indicator on the page measures outbound data in real time, so you can watch it stay at zero while you translate. Mainstream online translators upload your file; this one does not.

It depends on your browser. In recent Chrome on desktop, it uses the browser’s built-in Translator API, which runs a translation model on your device and downloads a small language pack for each language pair the first time you use it, managed entirely by your browser. Otherwise, if your browser has WebGPU, it uses Qwen2.5-1.5B-Instruct (a 1.5-billion-parameter instruction model, 4-bit quantized, run through MLC web-llm on your GPU in a Web Worker) prompted to translate. The page shows which one applies before you start, and both keep the document on your device.

Pressing Download fetches the weights (about 850 MB) from this site over your own connection, same-origin, never from a third party. Your browser stores them (via IndexedDB) and caches them, so it is a genuine one-time download that is shared with Summarize PDF and Ask PDF, so it never downloads twice. The built-in Translator path downloads only small per-language packs, handled by your browser. In every case only model data is fetched; your PDF is never part of the download.

You can translate from a wide range of source languages (including non-Latin scripts, which are read on your device). This first version writes the translated PDF with a standard font, which renders Western-European languages, so the target language is English, Spanish, French, German, Portuguese, Italian or Dutch for now. Target languages that need a different script (Chinese, Japanese, Korean, Arabic, Hindi, Russian and others) are coming in a version that embeds a Unicode font.

Yes. The translation is laid back over the original page: your images, tables, and page layout are preserved, and each block of text is translated where it sits, rather than rebuilt as a plain text document. A couple of honest limits: the original words are covered with the local background colour (sampled from the page), which is exact on plain backgrounds and very close on coloured cells; and a translation much longer than the original is shrunk to fit its space. The result reads as your document, in the new language.

A scanned PDF has no selectable text, so there is nothing to translate directly. Run OCR PDF first (also on-device) to add a text layer, then translate the result.

No. Once the model or language pack is on your device, translation runs entirely in your browser with no network, during or after. The only network event the feature can cause is the one-time model or language-pack download coming to you; your document never goes the other way.

It needs a desktop or laptop: either a browser with the built-in Translator API (recent Chrome on desktop) or a browser with WebGPU (recent Chrome, Edge, or other WebGPU-capable browsers) and a GPU with enough memory. Phones and tablets are not supported: the WebGPU model is a 1.5-billion-parameter AI (about 850 MB) that a mobile GPU cannot run, and loading it can exhaust memory and crash the device, so the tool does not offer it there. If a device cannot run it, the tool says so and does nothing; it never falls back to uploading your file. Every other PDF tool here still works regardless.

Not quite, and the difference is the one that matters. Some translators that call themselves private keep your file on your device, but they extract the text and send that text to a translation server, so the words in your contract or record still leave your machine. Here, both the file and the text stay on your device: the model runs in your browser, and the live 0-byte meter shows nothing goes out. That is the specific edge, even the text never leaves.

It depends on which on-device path your device uses. On a desktop with the built-in Translator (recent Chrome), you get the browser’s own on-device translation, which is high quality. On other devices, including most phones, it runs a compact model on your device instead: the translation is still useful and completely private, but it can read less fluently than a cloud service like DeepL or Google Translate. That trade is deliberate, quality that stays entirely on your device rather than quality bought by uploading your text.

To keep it fast and within the model’s context, very long documents are translated from their opening section, and the tool tells you when that happens. For a full-document translation of a long file, split it first and translate the parts.