Translate PDF without uploading it

Got a German contract draft or a Spanish user manual you need to understand before you reply? The translator pulls the text out of the PDF, groups it into paragraphs and translates it block by block. You can see which page it's on and how long is left, and you can stop at any point – whatever has been translated so far is kept.

You get the result three ways:

• Side by side: the original paragraph on the left, the translation on the right, page by page. It's the easiest way to spot a sentence that went sideways.
• Translated PDF: the original pages with the translated text placed roughly where the original text was. Images, lines, table borders and logos stay put.
• Word (DOCX) and plain text (TXT): for editing the translation or pasting it somewhere else.

The source language is detected automatically, but you can change it. If you only need a few pages, enter them (for example 1-3, 7) and it finishes faster.

Translation doesn't happen on a remote server – it happens on your device. The tool uses one of two engines and always tells you which one did the work.

	Built-in browser translator	Open translation model (OPUS-MT)
Where it runs	Desktop Chrome 138 or later	Any modern browser, including phones
Languages	Many language pairs	English, German, French, Spanish, Italian, Dutch, Russian, Ukrainian, Czech, Finnish, Swedish, Danish, Hungarian; Polish to English
First use	Chrome downloads a language pack	The model downloads (about 110–140 MB per direction)
Speed	Fast	Slower, typically 10–30 seconds per page

The built-in translator is Chrome's own on-device model. When your browser has it, the tool suggests it first, because it's faster and covers more languages.

The open translation model is OPUS-MT from the University of Helsinki, running in your browser on WebAssembly. It only downloads after you click, and the size is shown up front. After that it stays in your browser's cache, so the next translation starts without a download. When neither language is English (say, German to Spanish), the translation goes through English using two models. That means twice the download and slightly lower accuracy than a direct pair.

To build the translated PDF, the tool removes the original text from each page and writes the translation into the same box the paragraph occupied. Translations often run longer than the source – German and Hungarian are notorious for it – so the font size shrinks automatically until the text fits. Headings stay bold, and long paragraphs are wrapped for you.

This is a translation overlay, not a re-edit of the original document, so it has limits worth knowing:

• It uses one font (DejaVu, which covers accented Latin, Greek and Cyrillic), not the original typeface.
• Text is set in a dark color rather than the original colors.
• Text baked into images – charts, scanned stamps – isn't translated.
• If a paragraph sits in a very tight space, the translation can end up very small. The tool warns you when that happens.
• If the original text can't be removed (for example, text inside embedded form objects), the translation is placed on white cover boxes.
• For Japanese, Chinese and Korean targets, the PDF font doesn't include the characters. Use the Word or TXT download instead.

If you'd rather keep the original text in the file – to keep it searchable, for instance – turn off removing the original text. The translation then appears on white boxes over the original.

A scanned PDF is a set of page images with no selectable text. When the tool spots one, it offers text recognition (OCR): it reads the letters on each page (English and Hungarian language data), groups them into paragraphs and translates those. In the translated PDF, scanned pages get the translation on white boxes over the recognized paragraphs, since text can't be removed from an image.

Recognition is only as good as the scan. A skewed, blurry or handwritten page means recognition errors, and recognition errors mean translation errors, so compare the side-by-side view carefully.

Machine translation is great for understanding a foreign-language document quickly, deciding whether you need to deal with it, and pulling out the parts that matter. It isn't a certified translation: courts, government agencies and contract signings usually require a sworn or certified translator.

A few things to watch:

• Numbers, dates and names: check them against the original. That's what the side-by-side view is for.
• Short headings and table cells: a few words give the model little context, so they're translated less reliably than full sentences.
• Jargon: the open model was trained on general text and can miss a technical term. If accuracy matters and your browser supports it, try the built-in translator too.
• Long documents: with the open model, a 100-page document takes a while. If you only need a few chapters, enter the pages.

Keep the tab open while it works. Your file never leaves your device: the only network traffic is the model download (the open model from Hugging Face, its runtime from jsDelivr, language packs from Chrome itself). Those downloads bring the model to you; your document's text never goes out.