babel
LATAM · self-hosted LLM · glossary-locked

Translate a whole book without the 50-page cap.

Drop a PDF, EPUB, or DOCX. babel runs an open-source translation model on a real GPU, locks your proper nouns so names stay consistent across chapters, and emits a clean .docx or .epub. Spanish (Argentina), Spanish (Mexico), Portuguese (Brazil) and 37 other languages with proper regional flavor.

Start translating →Download for MacSee how it works
No page capsGlossary reviewes-AR · es-MX · es-ESConsistent names across chapters

Three steps, one coffee.

Most books finish before your third refill.

  1. 1

    Upload

    PDF, EPUB, DOCX, TXT, or Markdown. We parse the layout, count tokens, and detect the source language.

  2. 2

    Review the glossary

    babel pulls recurring proper nouns and lets you fix how each one should be translated — Alice → Alicia, not Alisha.

  3. 3

    Translate + download

    Chunks run sequentially with the previous passage as context. Pick your output format when it's done.

What the others don't do.

Regional Spanish that actually sounds regional

Pick rioplatense, Mexican, Peninsular, or LATAM-neutral. babel injects the right register — voseo, vocabulary, idioms — into every chunk's prompt.

Glossary lock

Review proper nouns before the long run. Your character names stay spelled the same way from page 1 to page 500.

No page or word caps

DeepL caps you at a handful of pages per month. BookTranslator stops at 50 MB. babel runs whatever fits on disk.

Real file formats, not plain text

Drop a DOCX with headings, get a DOCX back. EPUBs preserve chapters. Markdown in, Markdown out.

Your text stays yours

Dedicated infrastructure. No retraining, no third-party LLM APIs unless you opt in. Source and output files are deleted after the retention window.

Priced for LATAM

Free tier for small docs while the beta is open. Paid tiers coming soon in USD, ARS, and MXN.

Your manuscript, your terms.

Unreleased books don't belong on someone else's training set.

✓ What babel does

  • · Runs the translation model on dedicated infrastructure (currently our own GPU).
  • · Deletes your source and output files after the retention window.
  • · Lets you self-host the entire stack — code is open on GitHub.

✕ What babel doesn't

  • · Send your text to any third-party LLM API unless you explicitly opt in.
  • · Keep your documents indefinitely.
  • · Train a model on your uploads.

Frequently asked

What model does babel use?⌄
Google's TranslateGemma (4B and 12B variants), benchmarked on 55 language pairs. It runs locally via llama.cpp on Metal, CUDA, or CPU.
How long does a book take?⌄
A 300-page (~80k word) novel on a modern NVIDIA card finishes in 20–40 minutes. On an M-series Mac, about double that.
Can I really use it for free?⌄
During the beta, yes, subject to queue limits and file-size caps. Admin-mode users (authors with a pass-code) get unlimited throughput.
What if I want OpenRouter or OpenAI instead?⌄
Cloud adapters are stubbed — you can swap in your own API key. That route bypasses our local GPU and your text goes through the third-party provider.
Open source?⌄
Currently private during beta. Selected invited users can clone and self-host.

Ready to translate?

The first chapter takes about a minute. No account required to try.

Open babel →