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 capable of translating 200 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 specifically to cover languages that most commercial translation services underserve — alongside major world languages, it includes dozens of lower-resource languages that rarely get dedicated attention from big tech translation products.

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. That compression is precisely why the download exists at all — a model that understands 200 languages well enough to translate between arbitrary pairs of them simply isn't a small file, even after significant optimization.

How the Quality Honestly Compares

Being straightforward about this matters more than sounding impressive: for major, widely-spoken language pairs — Spanish to English, French to German, and similar — Google Translate and DeepL generally produce more fluent, natural-sounding output. Those services run considerably larger models on dedicated server hardware, without the constraint of fitting inside a browser tab.

Where the comparison shifts is lower-resource languages — ones that large commercial services often support only patchily, if at all. Because broad coverage was NLLB-200's specific design goal, it frequently produces more usable results for these languages than mainstream alternatives manage. The honest summary: for a major language pair, treat this as a strong, completely private option that may read slightly less polished than a cloud giant; for everything else, it may well be the better tool outright.

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.
• Lower-resource languages that mainstream translation tools handle poorly or not at all, where this model's broader coverage genuinely helps.
• Personal and family documents — letters, records, certificates — translated without a third-party service ever seeing them.

Frequently Asked Questions