PDF2zh (PDFMathTranslate): Scientific PDF Translation

When you need to read a paper in another language, a summary may not be enough. PDF2zh translates scientific PDFs while aiming to retain the layout around paragraphs and equations. You provide a PDF you are entitled to use; the intended result is a translated or bilingual document that you can compare with the original.

PDFMathTranslate, also called PDF2zh, is a third-party open-source project, not software developed by Gwarket. This page is based on public documentation and visible browser interfaces. We did not upload a document or test a translation. Both formatting and translation accuracy need checking.

Choose the output you need

These tools can work together. Translation supports reading; it does not establish the strength of the evidence. Check quotations against the original. Gwarket's evidence-led paper-reading guide can help you organize claims and source locations afterwards.

Inspect the upstream examples first

The project provides an animation comparing original pages with Chinese translations. This is the upstream project's demonstration, not a Gwarket translation test. We link to it rather than rehost its book or paper pages.

Look for three things in the demonstration:

  1. Do equations and their numbers remain near the relevant paragraphs? A preserved appearance still needs a content check.
  2. Were labels inside figures translated? Some architecture-diagram labels in the upstream demonstration remain in English.
  3. Can you compare columns and captions easily? Selected examples do not guarantee the same outcome for another PDF or the current hosted services.

A minimum useful workflow

  1. You provide: a short, text-based PDF that you can lawfully use and share publicly. Decide whether you need a translated document or only a summary.
  2. The tool processes: choose a hosted service, languages, and pages, then try to obtain a translation. A service may introduce additional terms or sign-in requirements.
  3. You check: compare the output with the original, inspect numbers and negation, then equations and layout. Check before quoting.

Start in a browser with a short public PDF

The original project README links to several hosted services. The two below have separate limits and data practices; one service's rules do not establish another's.

PDF2zh public interface

pdf2zh.com offers a file or link input, translation-service selection, and source and target language controls. On November 3, 2026, it displayed a file limit of less than 5 MB and a first-20-pages option.

  1. Choose a short, text-based PDF that you are allowed to use and share publicly.
  2. Select the document and translation service, then the source and target languages.
  3. Check the page selection and current terms before clicking Translate.
  4. If the service produces a result, save it and compare it with the original page by page.

These steps describe the visible controls, not a tested conversion. We did not find a clear, verifiable retention and deletion period for this interface. Avoid confidential files.

Immersive Translate's BabelDOC

On the same date, the BabelDOC service page listed 166 pages per translation for free users and a 500 MB file limit. Those are this service's limits, not a shared allowance for every PDF2zh deployment. Check the current page before use.

The service says scanned, image-based, and OCR-processed scanned PDFs are unsupported. An experimental OCR workaround is mentioned separately, so do not depend on scanned documents working. It also says chart areas are not translated and references are ignored by default.

Check where your document goes

The BabelDOC page says it collects uploaded file contents and translation error logs for product optimization. It keeps translation records and related documents free for 14 days; afterwards, they may be deleted, for example during cache clearing. This does not guarantee deletion on day 14, and it does not describe every hosted service or translation provider.

Unpublished papers, customer records, personal data, and confidential documents require authorization and an assessment of the applicable data practices before upload. Open-source software does not mean a hosted service runs locally on your computer or never stores files. If a service asks for sign-in or payment, review the terms before continuing.

Four checks before relying on a translation

  1. Numbers and negation: compare sample sizes, percentages, units, and the wording of conclusions.
  2. Equations and figures: check symbols, subscripts, captions, and tables. A preserved appearance does not prove accuracy.
  3. Layout and omissions: look for shifted or missing text in columns, footnotes, and paragraphs spanning pages.
  4. Terminology and quotations: check consistency and return to the source before quoting.

The project's EMNLP 2025 paper describes layout detection, translation, and re-rendering. That explains the design, not success on every PDF. We have not measured speed or queues or verified translation output.

Keep the original project separate from version 2.0

The original README directs version 2.0 to a separate PDFMathTranslate-next repository. The original repository had more than 38,000 GitHub stars on November 3, 2026. That count is not the separate 2.0 repository's count and does not guarantee quality. Hosted services and source-code versions do not necessarily change together.

The original project uses the AGPL-3.0 license. This page provides an introduction and upstream links; it does not redistribute the software or offer a Gwarket-hosted translation service.

Documentation and interfaces checked on November 3, 2026. No service sign-in, tool installation, model download, document upload, or translation test was performed.