A PDF translator that doesn't mangle your LaTeX

What it does

PDFMathTranslate takes academic PDFs and produces translated versions that actually look like the original. It spits out two files: a monolingual replacement and a side-by-side bilingual document. Formulas, charts, tables of contents, and annotations are supposed to survive the trip intact. It supports a laundry list of translation backends—Google, DeepL, OpenAI, Ollama, MiniMax, and others—and wraps them in a CLI, a browser GUI, a Windows executable, Docker, and even a Zotero plugin.

The interesting bit

Most PDF translators treat the document as a bag of text and let the layout die. This one uses a layout-analysis model (DocLayout-YOLO) to map the document structure before touching a word, which is why the math doesn’t turn into gibberish. The experimental v2.0 kernel adds an isolated “precise” mode if you want to sandbox the process further.

Key highlights

• Preserves formulas, charts, table of contents, and annotations instead of flattening everything to text
• Supports multiple translation services: Google, DeepL, OpenAI, Ollama, MiniMax, and more
• Multiple interfaces: CLI, browser-based GUI, Windows executable, Docker container, and Zotero plugin
• Generates both example-mono.pdf and example-dual.pdf for monolingual or side-by-side reading
• 35k+ stars and an EMNLP 2025 Demo badge, suggesting the academic crowd has voted with its GitHub stars

Caveats

• Local installation requires Python 3.11–3.12; anything outside that range is unsupported
• The layout model (wybxc/DocLayout-YOLO-DocStructBench-onnx) must download on first run, which can hang for users behind certain network blocks
• Public demo instances run on limited resources, so heavy use or large documents may hit walls

Verdict

Worth a look if you regularly read scientific papers in languages you don’t speak and need the original layout intact. Skip it if you’re just converting plain text or simple Word docs where formatting is disposable.