google/translategemma-12b-it

TranslateGemma: Lightweight Multilingual Translation

TranslateGemma is a family of 12 billion parameter open translation models developed by Google, built upon the Gemma 3 architecture. These models are specifically designed for efficient and accurate translation across 55 languages, making them highly versatile for global applications. A key differentiator is their ability to perform both direct text translation and text extraction and translation from images (normalized to 896x896 resolution).

Key Capabilities

• Multilingual Translation: Supports translation for 55 languages, utilizing ISO 639-1 Alpha-2 codes or regionalized variants.
• Image-to-Text Translation: Can extract text from images and translate it, offering a powerful feature for diverse use cases.
• Resource-Efficient Deployment: Its lightweight nature allows for deployment on devices with limited resources, such as laptops, desktops, or private cloud infrastructure.
• Flexible Input Handling: Features a specific chat template that supports both text and image inputs, with a total input context of 2K tokens.

Good For

• On-device Translation: Ideal for applications requiring translation capabilities in environments with constrained computational resources.
• Multilingual Content Processing: Excellent for translating text across a wide array of languages.
• Visual Text Translation: Suitable for scenarios where text needs to be extracted and translated from images, such as signs or documents.
• Innovation in Translation: Fosters innovation by providing accessible, state-of-the-art translation models for developers and researchers.