Translate PDF

No PDF comes back: you get Markdown and plain text, without the original's fonts, columns, images or page positions. To typeset it, paste it into your own document or run translation.md through Markdown to PDF. That tool refuses scripts it cannot draw correctly, so Arabic or Thai output has to go into your own document. It also refuses a translation.md carrying any page notice, because the ⚠️ in each notice counts as emoji there; delete that sign first.

Not directly. Only text the PDF already contains as text is read, and your browser stops a file with none before it uploads or draws on the free allowance. Run the scan through OCR PDF with the right recognition language and the Searchable PDF output, then translate the PDF it returns.

The provider receives the extracted text and any instruction you typed, not the PDF file, under agreements that prohibit training on data sent through the API. pdf.xyz cannot control what a provider does beyond those terms, so if they are not enough for a document, keep it out of every AI tool.

A page over 10,258 characters cannot fit one pass, so it is skipped rather than cut off, and a notice where it sat names the page and its character count. Extracting that page on its own does not help, because the page itself is over the limit. Separately, if our server finds more text than the job reserved from the free allowance, or more pages with text than your browser counted, translation stops at that page and one notice there names the untranslated pages and which limit ran out.

It is machine output and may contain errors, as the package's README says. The model is told to leave an unintelligible passage as it is rather than guess, and the manifest names the model. Check important passages against the original.

Up to 100 MB per file.

Yes. 30,774 free characters of AI text a day for your connection — the allowance resets daily. The other AI tools draw on the same allowance.

10,258 characters of selectable text per pass — about 5 sparsely-set pages, or 1 page of dense text. Translation writes roughly as much as it reads, so that ceiling is the model's output budget — far below the reading cap the other AI tools state. A longer document is not refused for length: it is split at page boundaries, translated a few pages at a time, then reassembled in order. A SINGLE page holding more than that many characters cannot be translated at all — it is skipped, and the result says so in place, at the point where that page was. The whole document is charged against the free daily allowance when the job is created, so a document needing more than is left today is refused before it runs, with nothing charged, rather than half translated; one over 30,774 characters, the whole day's allowance, has to be split into smaller files. If our server finds more text or more pages than the job reserved, translation stops there and a notice in place names the untranslated pages and why, so nothing stops silently.

21 languages: English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Ukrainian, Polish, Swedish, Turkish, Arabic, Hindi, Chinese (Simplified), Japanese, Korean, Vietnamese, Indonesian, Thai and Bengali. The source language is detected automatically (“Auto-detect”), or you can pick it from the same list.

Yes. The result is written by an AI model from the PDF's selectable text, and a model can miss or misread details. Check anything you rely on against the document itself.