TranslateGemma - A series of open-source translation models from Google.
TranslateGemma is a new generation of open-source translation models from Google, based on Gemma 3. The model supports high-quality translation of 55 languages through a combination of supervised fine-tuning and reinforcement learning, significantly improving translation quality...
What is TranslateGemma?
TranslateGemma is a new generation of open-source translation models from Google, based on Gemma 3. The model supports high-quality translation of 55 languages through a combination of supervised fine-tuning and reinforcement learning, significantly improving translation accuracy and efficiency. TranslateGemma offers models with three parameter scales: 4B, 12B, and 27B, suitable for mobile devices, consumer laptops, and high-performance cloud computing environments, respectively. The model excels in text translation, retaining strong multimodal capabilities and translating text in images without requiring dedicated multimodal training.
The main functions of TranslateGemma
-
Multilingual translationThe model supports high-quality translation of 55 languages, covering high-resource languages (such as English, Chinese, and French) and low-resource languages (such as Icelandic and Swahili).
-
High performanceThrough model optimization, the smaller 12B model outperformed the baseline 27B model, and the 4B model provides powerful translation capabilities on mobile devices.
-
Multimodal translationThe model retains the multimodal capabilities of Gemma 3, enabling it to translate text in images without requiring additional multimodal training.
-
Wide applicabilityThe model is suitable for a variety of deployment environments, including mobile devices, consumer laptops, and cloud-based high-performance computing platforms.
The technical principles of TranslateGemma
-
Supervised Fine-Tuning (SFT):The model was fine-tuned using high-quality parallel data, including human-translated text and synthetic translations generated by the Gemini model.In this way, the model significantly improves translation quality in low-resource languages while maintaining its general capabilities.
-
Reinforcement Learning (RL):The model optimizes translation quality using various reward models, such as MetricX-QE and AutoMQM.Through reinforcement learning, the model can generate more natural and accurate translation results, further improving translation quality.
TranslateGemma's project address
- Project official website: https://blog.google/innovation-and-ai/technology/developers-tools/translategemma/
- HuggingFace model libraryhttps://huggingface.co/collections/google/translategemma
- arXiv technical paper: https://arxiv.org/pdf/2601.09012
Application scenarios of TranslateGemma
-
Personal communication and travelThe model helps individuals overcome language barriers in instant messaging and travel, translating chat content or menus, road signs, etc. in real time.
-
Content creation and distributionThe news media and film and television industries can quickly translate content into multiple languages, expand their international influence, and improve the global accessibility of their content.
-
Enterprise internationalizationBusinesses use translation tools to provide multilingual customer support, optimize marketing copy, improve cross-border collaboration efficiency, and facilitate globalization.
-
EducationThe model can assist students in learning foreign languages, help researchers quickly translate foreign language documents, and promote language learning and academic exchange.
-
Multimodal applicationsAutomatically translate text in image and video content, such as translating road signs or video captions, to improve the comprehensibility of multimodal content.