Translation Engine

Harvey provides two distinct model families for local translation. You can switch between them in the Configure tab under the Translation section.

Note: Only models belonging to the currently selected model family will be available for translation. If you switch families, you must download at least one compatible model for that family.

1. Helsinki-NLP

• Pros: Lightweight and very fast on CPUs. Provides high quality for common language pairs.
• Cons: Requires downloading a separate model for every language pair (e.g., Japanese to English, French to English).
• Recommendation: Best for users who primarily work with a few specific, common languages and value speed on standard hardware.

2. NLLB (Meta)

• Pros: A single model supports over 200 languages. Excellent for rare or less common languages.
• Cons: Much larger file sizes and slower on lower-end CPUs. Performance is best with a dedicated GPU.
• Recommendation: Best for researchers working with a wide variety of languages or rare dialects where specific Helsinki-NLP pairs may not exist.

CTranslate2 Backend

Both engines leverage the CTranslate2 backend for optimized inference. If prompted, ensure you install the required dependencies via the “Install Now” button in the Configure tab to enable high-speed translation.