TranslateGemma

TranslateGemma is a new family of open translation models built on Gemma 3, designed for developers and researchers to achieve high-fidelity translation quality with less than half the parameters of the baseline model. The models are available in 4B, 12B, and 27B parameter sizes, making them suitable for diverse deployment environments. TranslateGemma has been rigorously trained and evaluated on 55 language pairs, ensuring reliable, high-quality performance across major languages and low-resource languages. TranslateGemma sets a new standard for open translation models, balancing state-of-the-art performance with exceptional efficiency. The models retain the strong multimodal capabilities of Gemma 3, translating text within images, and have been tested on the Vistra image translation benchmark. The efficiency of these models allows for higher throughput and lower latency without sacrificing accuracy, making them ideal for mobile inference and edge deployment.