RepoPicks Radar

Updated every day around 23:30 UTC

Guides

PDFMathTranslate

PDFMathTranslate/PDFMathTranslateReviewed by AI

Checked by RepoPicks Radar on

What is PDFMathTranslate?

PDFMathTranslate (pdf2zh) translates PDF documents into another language and keeps the original layout: formulas, charts, tables, contents and notes stay in place, and it can also produce a bilingual file with both languages. It is built for scientific papers. On Windows you unzip the release and double-click pdf2zh.exe; on Mac and Linux you install it with pip or run the Docker image; the authors also keep free online pages. Translation goes through Google by default, or DeepL, OpenAI, Azure and local Ollama models.

PDFMathTranslate is free and open source (AGPL-3.0) and can replace DeepL Pro, Adobe Acrobat, and Mathpix.

How easy is it to install PDFMathTranslate?

A · download and open · Windows

  • Windows A
  • macOS C
  • Linux C
  • Web A

Windows and the web share grade A. On Windows you "Download pdf2zh-version-win64.zip from release page" (308 MB, or 521 MB with assets), "Unzip and double-click `pdf2zh.exe` to run", with everything bundled; the README only warns that some machines first need Microsoft's vc_redist.x64.exe. On the web the authors run a free service at pdf2zh.com plus Hugging Face and ModelScope pages. macOS and Linux are C: there is no ready file for them, so you install Python 3.11 or 3.12, then `pip install pdf2zh` and run `pdf2zh -i` to open the page at http://localhost:7860/.

PDFMathTranslate on GitHub

Stars
38,349on Nov 10, 2026
Downloads
147,9151,403 this week
Growth
+179stars in 7 days

Stars over time

38,20038,25038,30038,350 Nov 5Nov 6Nov 7Nov 8Nov 9Nov 10 38,20038,25038,30038,350 Nov 5Nov 8

Facts

Runs on
Windows, macOS, Linux, Web
You need to install
Nothing extra
To build from source
Python, Docker, NVIDIA graphics card
Disk space
521 MB–1.5 GB · from the installer
Accounts or keys
OpenAI (only if you choose it as the translation service), DeepL, Azure Translator, Tencent Machine Translation
Cost
Free · The program is free (AGPL-3.0) and the default Google service needs no account. Picking DeepL, OpenAI, Azure or Tencent instead means paying that service for what you translate.
License
AGPL-3.0 (Copyleft)
Can replace
Pro, Acrobat, Mathpix
Last change on GitHub
Nov 10, 2026

PDFMathTranslate risk signals (8)

  • Attentionyour document is sent to a translation service

    README section 4.1: "Use Google as the default translation service", and the dependencies include googletrans, deepl, openai, azure-ai-translation-text and tencentcloud-sdk-python-tmt. So by default the text of your PDF travels to Google, and only the ollama and xinference-client options keep it on your own machine.

    AI review

  • Attentionone flag puts the page on the public internet

    The options table lists `--share` with the function "Public link" (`pdf2zh -i --share`), which is the Gradio feature that hands out an address anyone can open, while sign-in is a separate opt-in (`--authorized users.txt`). The Docker instructions also publish the interface with `docker run -d -p 7860:7860 byaidu/pdf2zh`.

    AI review

  • Attentionthe ready windows file is from july 2025

    The newest release is v1.9.11 from 2025-07-11 while the last commit is from 2026-11-08 and pyproject.toml already says version 1.9.12. The same pyproject carries the comment "# >=6.16.0 fixes the gradio 5.x advisories (e.g. CVE-2026-28414)" next to `gradio>=6.16.0`, a fix that cannot be inside a build packaged sixteen months earlier.

    AI review

  • Infoversion 2 lives in another repository

    README: "2.0 Moved to a new repository under the organization: PDFMathTranslate/PDFMathTranslate-next" and "Version 2.0 official release has been published", while section 4.3 says this repository is "the present and the original project for stable release" and the other fork "is intended solely for development and does not address compatibility issues and is not designed for community-contributions". Two active projects share the name.

    AI review

  • Infodownloads a layout model and ocr data on first use

    README: "The current program relies on the AI model (`wybxc/DocLayout-YOLO-DocStructBench-onnx`)", and in OCR mode "The first scanned page downloads the requested language data from Tesseract's `tessdata_fast` 4.1.0 release into `~/.cache/pdf2zh/tessdata/4.1.0`". The Windows release comes in two sizes, 308 MB or 521 MB "with-assets", which is the same difference.

    AI review

  • Infoa mirror is suggested for the model download

    For "network difficulties when loading the AI model" the README suggests `set HF_ENDPOINT=https://hf-mirror.com`, which fetches the model from a third-party mirror instead of Hugging Face itself.

    AI review

  • Infosponsors in the readme and rewards for contributors

    The README opens with a Warp sponsorship banner linking to go.warp.dev/PDFMathTranslate, and the Acknowledgement says "Immersive Translation sponsors monthly Pro membership redemption codes for active contributors to this project". Immersive Translate's BabelDOC is also one of the recommended online demos and the project's "new backend".

    AI review

  • Infoexperimental modes with stated limits

    The README marks OCR, `--mode precise` and `--ultrafast` as experimental and lists the limits itself: "The initial implementation targets white-background scans: partial scans on pages that already contain text are skipped, and handwritten text or inline equations may be recognized incorrectly", and ultrafast "is best suited to PDFs that already contain native text; use normal mode for complex tables or formulas".

    AI review

Evidence is shown in its original language.

What Radar could not check about PDFMathTranslate

  • How good the translation reads, since the wording comes from whichever outside service you pick
  • Whether the layout really survives on your own PDFs, especially scans, complex tables and inline formulas
  • What each translation service keeps of the text you send it
  • Real disk use beyond the 521 MB to 1.5 GB measured from the installers, once models and OCR data are cached
  • Whether the 2025 Windows build still behaves like the current code, which has moved on by sixteen months

Similar projects

The RepoPicks app

Coming soon: install in 1 click with the RepoPicks app.

Before you download

See the release page on GitHub