google/translategemma-27b-it
TranslateGemma-27b-it is a 27 billion parameter open translation model developed by Google Translate, based on the Gemma 3 family. Designed for translation across 55 languages, this lightweight model supports both text-to-text and image-to-text translation, making it suitable for resource-constrained environments. It excels at direct translation and text extraction from images, offering high-performance translation capabilities.
Loading preview...
TranslateGemma-27b-it: Lightweight Multilingual Translation
TranslateGemma-27b-it is a 27 billion parameter model from Google Translate, part of the Gemma 3 family, specifically engineered for translation tasks. This model is designed to be lightweight, enabling deployment in environments with limited resources such as laptops, desktops, or private cloud infrastructure, thereby democratizing access to advanced translation technology.
Key Capabilities
- Multilingual Translation: Supports translation across 55 languages.
- Text-to-Text Translation: Directly translates text inputs from a source to a target language.
- Image-to-Text Translation: Capable of extracting text from images (normalized to 896x896 resolution) and then translating it.
- Resource-Efficient: Its relatively small size allows for deployment in resource-constrained settings.
Usage and Performance
The model utilizes a specific chat template for direct translation or text extraction and translation from images. It supports ISO 639-1 Alpha-2 language codes and regionalized variants. Evaluation metrics show strong performance, with a COMET score of 84.4 and an MQM score of 5.86 on WMT25 (10 langs), demonstrating superior performance compared to other comparably-sized open models. Training involved 4.3 billion tokens during supervised fine-tuning and 10.2 million tokens during reinforcement learning, using a mix of monolingual web documents paired with Gemini-generated translations and publicly available parallel documents.