TranslateGemma: A new suite of open translation models (blog.google)

Today marks the launch of TranslateGemma, an innovative suite of open translation models built on the Gemma 3 architecture, available in 4B, 12B, and 27B parameter sizes. This development is significant for the AI/ML community as it enhances communication capability across 55 languages while emphasizing efficiency. Notably, the 12B model outperforms the larger 27B baseline model, demonstrating a remarkable advancement in translation quality achieved through a specialized two-stage fine-tuning process. This breakthrough allows developers to maintain high-quality translations with fewer parameters, ultimately increasing throughput and reducing latency in deployment. TranslateGemma's training incorporated a diverse dataset, including both human translations and synthetic outputs from Gemini models, ensuring high fidelity even for low-resource languages. Its multimodal capabilities extend to translating text within images, enhancing its usability across various applications. With models tailored for diverse environments—from mobile devices to cloud infrastructure—the suite is poised to empower researchers and developers to refine language translation technologies. The release sets a new benchmark in open translation, encouraging further exploration and adaptation within the AI community to facilitate greater understanding across cultures.