Google said evaluations indicated that TranslateGemmas 12B model outperformed the larger Gemma 3 27B benchmark model on WMT24++, while requiring significantly fewer parameters to do so.