Translate a PDF

Translation is a different category of problem from everything else this site does. Compressing, splitting, or cropping a PDF is a matter of following well-defined rules — there's a correct answer, and a program can compute it directly. Translation isn't rule-based in the same way; it requires something that has actually learned the patterns of two languages well enough to produce fluent, meaningful output. That "something" is a neural network, and neural networks broad enough to translate reliably between many languages are large by nature.

This guide explains what that model actually is, why it downloads once rather than living permanently on a server somewhere, how its quality genuinely compares to the cloud services you've probably used before, and why a translated PDF ends up laid out differently from the original rather than an exact visual copy.

The Model: Meta's NLLB-200

This tool runs NLLB-200 ("No Language Left Behind"), an open-source translation model published by Meta AI that can translate between roughly 200 languages. To keep this page fast and the dropdown easy to scan, it's intentionally narrowed down to the 10 languages people most often need for a PDF: English, Mandarin Chinese, Hindi, Spanish, Arabic, French, Bengali, Portuguese, Russian, and Indonesian — together the ten most-spoken languages in the world.

The version used here is a distilled variant: a smaller model trained to approximate the behavior of Meta's full-sized NLLB-200, compressed enough to run inside a browser tab using WebAssembly, without needing a GPU or a server. On top of that, it's loaded in a quantized (int8) form, which shrinks the download to roughly 900MB instead of the 3.5GB the uncompressed version would need, with no meaningful difference in translation quality for this particular model.

How the Quality Honestly Compares

Being straightforward about this matters more than sounding impressive: all 10 languages this tool offers are also the ones Google Translate and DeepL invest the most in, so for any pair among them, those services will generally produce a shade more fluent, natural-sounding output. They run considerably larger models on dedicated server hardware, without the constraint of fitting inside a browser tab.

What this tool trades for that small quality gap is privacy: your document's content never leaves your device, while Google Translate and DeepL both send your text to a server to be translated. For anything sensitive — legal paperwork, medical records, personal correspondence — that trade-off is usually worth a translation that reads very slightly less polished.

Why the Layout Changes

Translated text is almost never the same length as the original — German sentences commonly run longer than their English equivalents; Chinese text is often visually denser than the same meaning in Spanish. Fitting translated text back into the exact pixel positions the original occupied would mean overflowing text boxes, overlapping lines, or awkward automatic shrinking, none of which produce a genuinely readable document.

Rather than attempting a fragile visual match, this tool extracts the text in its original reading order, translates it, and lays it out fresh on clean pages — proper paragraph spacing, sensible pagination, nothing overlapping. The result reads naturally in the new language, which is a more useful outcome than a pixel-accurate copy that's actually hard to read.

Why "Nothing Uploaded" Matters More for Translation

Documents people need translated are very often the ones they'd least want passing through an unfamiliar server: immigration paperwork, medical records, contracts, personal correspondence. Nearly every PDF translator online works by sending your text to a translation API somewhere — that's simply how cloud translation services are built, DeepL and Google Translate included.

Because the entire model runs on your own device after that first download, your document's content is never transmitted anywhere to be translated. You can verify this directly: open your browser's developer tools, watch the Network tab during translation, and confirm that beyond the one-time model download, nothing carrying your document's text goes out.

Common Reasons to Translate a PDF

• Immigration and legal paperwork, where the content is sensitive and a private, on-device option is worth the trade-off against maximum polish.
• Academic papers in a language outside your own, translated for a first-pass understanding before deeper reading.
• International business correspondence, where confidentiality matters as much as getting the gist across quickly.
• Diaspora and multilingual families, translating between English and a home language like Hindi, Bengali, Mandarin, or Arabic to share documents across generations.
• Personal and family documents — letters, records, certificates — translated without a third-party service ever seeing them.

Frequently Asked Questions