TranslateGema

On January 16th, Google released a book yesterday, January 15, which, based on the Gemma 3 architecture, launched the TranslateGema Open Translation Model series, with a total of 4B, 12B and 27B parameter sizes, supporting 55 core languages and multiple mode image translations, which are now available for download in Kaggle and Hugging Face. In terms of performance, Google teams use the WMT24++ benchmark (55 languages in high, medium and low resource languages) and MetricX..