About TranslateGemma
TranslateGemma is a new suite of open translation models, built upon the Gemma 3 architecture, available in 4B, 12B, and 27B parameter sizes. These models are designed to facilitate communication across 55 languages, offering high-fidelity translation quality with remarkable efficiency. The 12B model, for instance, outperforms the Gemma 3 27B baseline using less than half the parameters, making it ideal for developers seeking high throughput and low latency without compromising accuracy. The 4B model is also powerful for mobile inference.
These models are developed using a specialized two-stage fine-tuning process, distilling knowledge from Gemini models. This includes Supervised Fine-Tuning on diverse parallel data and Reinforcement Learning using advanced reward models to ensure contextually accurate and natural-sounding translations. TranslateGemma also retains the strong multimodal capabilities of Gemma 3, allowing for translation of text within images. It is available for download on Kaggle and Hugging Face, and can be deployed in Vertex AI, making it a robust foundation for researchers and developers to adapt for specific language pairs or to improve low-resource language translation.