A translation engine that doesn't phone home

What it does

Argos Translate is a Python library for offline neural machine translation, built on OpenNMT and CTranslate2. It downloads language models as .argosmodel zip archives, supports GPU acceleration via an environment variable, and powers the better-known LibreTranslate API and web app.

The interesting bit

The automatic pivoting is the quietly clever part: if you have Spanish→English and English→French models installed, Argos will chain them to translate Spanish→French even without a direct model. The README is upfront that this costs translation quality, which is rarer honesty than you’d expect.

Key highlights

• Runs fully offline once models are downloaded
• ~40 languages supported via package index (Arabic to Urdu, Esperanto, Kyrgyz)
• Same engine under the hood as LibreTranslate’s hosted API
• GPU support through CTranslate2 with ARGOS_DEVICE_TYPE toggle
• HTML and file translation handled by companion libraries

Caveats

• The core project is a library/CLI; GUI lives in a separate repo
• No quality metrics or benchmarks published in the README
• Custom model training is a paid service ($1000/language, negotiable)

Verdict

Worth a look if you need embeddable, privacy-preserving translation without cloud dependencies. Skip it if you require top-tier accuracy or a polished out-of-box GUI — this is infrastructure, not a finished product.