PDFMathTranslate
An open-source PDF translator that preserves formulas and layouts, runs locally or in Docker, and supports Ollama, Google, DeepL and OpenAI.
PDFMathTranslate translates scientific PDFs while keeping their page layout, formulas, charts, contents pages and annotations. It's for researchers, students and others who need to read papers in another language without losing the relationship between the text and its figures. It produces both translated PDFs and bilingual documents for comparison with the original.
You can run it locally through a browser interface or command-line tool, use the Windows application, or self-host it with Docker. The Python project is open source under AGPL-3.0. Hosted demos are also available through HuggingFace and ModelScope; those run on external services rather than your own machine.
Translation backends include Ollama for local models, plus Google, DeepL, OpenAI and MiniMax. The backend choice determines where translation happens: Ollama supports local processing, while external translation services receive the text they translate. Document layout detection uses DocLayout-YOLO, and the tool needs to download its layout model unless it's already available locally.
Beyond whole-document translation, it supports selected pages, batches of PDFs and multiple source and target languages. It can work with local documents or PDFs available online, including papers on arXiv. A Zotero plugin connects it to a reference library, while Python and HTTP APIs let other applications use its translation functions. MCP support makes it accessible to compatible AI tools. The official README directs users of version 2.0 to the separate PDFMathTranslate-next repository.
