Google has launched TranslateGemma, a family of open-source AI translation models that support 55 languages and work efficiently across smartphones, laptops and cloud systems. Available in 4B, 12B and 27B parameter sizes, the models are designed to deliver high-quality translation with lower computing requirements.
The 12B model outperforms earlier larger models while using less compute, enabling powerful local translation, while the 4B model supports offline translation on low-power devices. Built by fine-tuning Gemma 3 with human and synthetic data and enhanced through reinforcement learning, TranslateGemma offers improved fluency and accuracy, including for low-resource languages.
With added multimodal capabilities to translate text within images, and open availability on platforms like Kaggle and Hugging Face, TranslateGemma advances accessible and efficient AI-powered multilingual communication.
