* The Decoder Article
Google's new open TranslateGemma models bring translation for 55 languages to laptops and phones
https://the-decoder.com/googles-new-open-translategemma-models-bring-translation-for-55-languages-to-laptops-and-phones/
# Part of this news was translated using the models mentioned.
Google's new TranslateGemma models demonstrate that smaller size can lead to better performance. A model with 12 billion parameters outperformed a model twice its size.
Google released TranslateGemma in three sizes: a 4 billion parameter model optimized for mobile devices, a 12B model designed for general consumer laptops, and a 27B model for cloud servers running on a single H100 GPU or TPU. … …
These models retain the multimodal capabilities of Gemma 3. They can translate text within images without any specific training. Vistra benchmark tests confirmed that improvements in text translation performance also apply to image-based translation.
For best results, Google recommends prompting the model as a "professional translator" to consider cultural context. The models are available on Kaggle and Hugging Face. … …
TranslateGemma is the latest addition to Google's growing Gemma product line. It is based on Gemma 3, an open-weight model for consumer hardware that Google launched in spring 2025. Since then, Google has developed specialized models for various use cases.
* Link to the Google Korea blog post.
‘TranslateGemma’: A new open-source translation model product line
https://blog.google/intl/ko-kr/company-news/technology/translategemma/
* Screenshot of using TranslateGemma 12B through LM Studio.
