Ratica

Offline AI PDF translator that keeps the original page layout

Translate PDF books on your own computer.
Free, open source and private: no cloud, no account, no API key. Windows · macOS · Linux.

Website · Download · Watch the 30-second intro · How to use · How it works

Ratica is an offline PDF book translator. It reads a PDF, translates it paragraph by paragraph with a local AI model (Gemma running on llama.cpp), and writes the translation back into the original pages: headings, lists, tables of contents, figures, colours and formulas stay where they were. You also get an EPUB for e-readers. Your book never leaves your machine.

Status: beta. Ratica works end to end, but it is young: expect rough edges and please report them.

Why Ratica?

Most PDF translators are cloud services: you upload your document, wait, and often get back plain text that has lost its layout. Ratica is a local PDF translator instead. It runs entirely on your own computer with a local AI model, so it works offline and your files stay private.

• Layout-preserving PDF translation: the translated book looks like the original, page by page.
• Private by design: no cloud, no account, no API key, no file uploads.
• Made for long documents: books, textbooks, manuals and technical documents of hundreds of pages; you can pause and resume.
• PDF and EPUB output: read the translation as the original pages or on an e-reader.
• Free and open source on Windows, macOS and Linux.

Translate between English, Turkish, Spanish, German, French, Japanese, Chinese and 20+ other languages, in any direction. Ratica recognises the language of the book by itself.

Install

Download the file for your system from Releases:

System	File	Notes
Windows 10/11	Ratica-…-Windows-Setup.exe	Run it. The installer is not code-signed yet, so Windows may show "Windows protected your PC": click More info → Run anyway.
macOS (Apple Silicon)	Ratica-…-macOS-AppleSilicon.dmg	Open the disk image, drag Ratica onto Applications and open it from there. The app is signed and notarized by Apple.
Linux (x64)	Ratica-…-Linux-x64.tar.gz	Unpack and run Ratica/Ratica.

On first start Ratica downloads its translation engine and AI model (about 5–6 GB, once) and measures your computer's speed. After that it works offline.

Run from source

Download or clone this repository, then double-click start.bat (Windows) or start.command (macOS), or run ./start.sh (Linux). The script installs uv if needed, then Ratica's packages, then opens the app.

Use

1. Choose a PDF or drop it on the window.
2. Pick the language to translate to. The book's own language is recognised for you; change it if needed.
3. Press Translate. You can pause and resume at any time; progress is saved. Want to try it first? Use the one-page sample book.
4. The translated book.<language>.pdf (same layout as the original) and book.<language>.epub are saved next to the original.

There is also a command line:

ratica translate book.pdf --to de        # translate (sets itself up on first use)
ratica translate book.pdf --to ja --keep "Kubernetes,Linus Torvalds"
ratica setup                             # download and measure speed now
ratica info                              # hardware and settings, for bug reports
ratica translate book.pdf --to fr --reflow   # typeset a new PDF instead of keeping the layout

Features

• Translates text-based PDF books (textbooks, manuals, papers, novels) in any direction between 23 languages: Azerbaijani, Bulgarian, Chinese (Simplified), Czech, Dutch, English, French, German, Greek, Hindi, Indonesian, Italian, Japanese, Korean, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Turkish, Ukrainian and Vietnamese.
• Recognises the book's language by itself; you only choose the language to translate to.
• Keeps the original layout: every page stays where it was, with its images, tables, charts, colours and page numbers. Each paragraph is translated in its own box, in a matching serif or sans-serif font; longer translations use free space next to them before the text gets smaller.
• Leaves code, formulas and URLs exactly as they are.
• Also writes an EPUB, which reflows nicely on phones and e-readers.
• Resumes where it stopped if the computer sleeps or the app closes.
• Shows a time estimate before starting.
• Sets itself up: it detects your hardware, downloads the right engine and model once, and measures your computer's speed.
• Leaves room for your other work: it never fills the GPU memory and runs at low priority.

How fast is it?

On an entry-level gaming GPU with 6 GB of memory, a 200-page book takes about 30–45 minutes. Faster GPUs finish sooner because they translate more paragraphs at once. Without a GPU, translation works but takes many hours.

Requirements

Hardware	Support
NVIDIA GPU with 6 GB or more	✅ Recommended (CUDA)
Apple Silicon Mac (M1 or newer)	✅ Supported (Metal)
AMD or Intel GPU	🧪 Should work (Vulkan), not yet tested — reports welcome
No GPU	⚠️ Works, but slow

Windows 10/11, macOS 12 or newer, and Linux. About 6 GB of free disk space for the engine and the model.

The AI model

Ratica uses Gemma-4-E4B-it (Apache 2.0) through llama.cpp. It was chosen after comparing several current open models on translation quality, protection of code and URLs, speed and memory use. The benchmark scripts are in bench/, so anyone can run the comparison on their own hardware.

Uninstall

The engine and the AI model (5–6 GB) are kept outside the app, in Ratica's data folder. To free that space while keeping the app, click delete next to Downloaded engine and model at the bottom of the window.

• Windows: uninstall Ratica from Settings → Apps. The uninstaller asks whether to delete the downloaded engine and model too.
• macOS: delete the downloaded files in the app first (or remove ~/Library/Application Support/Ratica), then move Ratica from Applications to the Trash.
• Linux: delete the Ratica folder and ~/.local/share/ratica.

Translated books and their saved progress (book.<language>.ratica folders) stay next to your PDFs; delete them like any other file.

Privacy

Everything runs locally. Ratica only goes online to download the engine and the model on first launch. It never uploads your files or text.

Known limitations

• Scanned PDFs (pages that are images) are not translated yet; OCR is planned.
• Text inside images stays in the original language.
• Formulas are kept exactly as they are, so a sentence that is part of a formula stays in the original language.
• Pages rotated sideways (wide tables) stay in the original language.
• When a paragraph mixes styles (an italic phrase, a coloured link), the translation uses the paragraph's main style for all of it.
• If a translation cannot fit its place on the page at a readable size, that text keeps the original.
• The app window is in English for now.

Roadmap

v1 (now)	v2
Text PDFs → PDF + EPUB	OCR for scanned pages
Code, URL, formula and name protection	Translated diagram labels
Resumable, parallel translation	Splitting large tables
Automatic hardware setup	Glossary for consistent terms
Simple desktop app	Side-by-side preview, translated app window

Right-to-left scripts and comics (text inside speech bubbles) are ideas for later, not promises.

Read the design in docs/ARCHITECTURE.md.

Contributing

Ratica is built in the open and help is welcome, especially:

• Hardware reports from AMD, Intel and Apple GPUs: speed, memory use, anything that breaks.
• Translation quality feedback in any language.
• Bug reports and pull requests.

Acknowledgements

llama.cpp · Gemma · PyMuPDF · Noto fonts · FLORES-200 · PySide6

License

AGPL-3.0. Ratica is free and open source. Anyone may use, study and change it; modified versions must stay open source too.