Open translation on Google models, supporting 55 languages
TranslateGemma is a new suite of open AI translation models built on Google’s Gemma 3. It enables high-quality communication across 55 languages, combining strong accuracy with exceptional efficiency. Designed to run on mobile, local devices, and cloud environments without compromising performance.
TranslateGemma supports 55 rigorously evaluated languages plus nearly 500 additional language pairs, with training on both human-translated and high-quality synthetic data from Gemini models, enabling reliable translation even for underrepresented languages.
TranslateGemma retains Gemma 3's multimodal capabilities, allowing developers to translate text embedded in images without separate fine-tuning, improving accessibility for visual content across languages.
The 12B TranslateGemma model outperforms Gemma 3 27B baseline using less than half the parameters, allowing researchers and developers to achieve state-of-the-art translation quality on consumer laptops and standard cloud infrastructure.
TranslateGemma is designed as a robust foundation for further adaptation, enabling researchers to fine-tune models for custom language pairs or improve quality for low-resource languages using their own parallel data.
TranslateGemma's 4B model is optimized for mobile and edge deployment, enabling developers to run high-quality translation locally without cloud dependency, reducing latency and improving privacy for end users.