Google translategemma 4B Translation Models (huggingface.co)

Google has launched TranslateGemma, a new family of lightweight, state-of-the-art translation models designed to democratize access to translation technology across 55 languages. Available on Hugging Face, these models accommodate both text and image inputs, enabling users to convert written content or extract and translate text from images. The TranslateGemma models are built on the existing Gemma 3 architecture and are optimized for deployment in resource-constrained environments, expanding the potential for innovative applications in translation on devices like laptops and cloud infrastructures. What sets TranslateGemma apart is its emphasis on user-friendliness and efficiency, utilizing a specific chat template compatible with Hugging Face's transformers. This allows for direct and organized communication between the user and the assistant roles, requiring clear parameters for language codes and content types. The training of these models leveraged Google's advanced TPU infrastructure, emphasizing performance, scalability, and cost-effectiveness. With a focus on ethical considerations, Google has conducted rigorous evaluations to minimize biases and enhance safety in translation outputs. As the AI/ML community looks to improve global communication tools, TranslateGemma represents a significant step forward in making advanced translation capabilities accessible to a broader audience.