TranslateGemma is a new suite of open translation models built on Google's Gemma 3 architecture, designed to enable high-quality communication across 55 languages. These models represent a significant step forward in open translation technology, helping people communicate regardless of their location or device.

The models are available in three parameter sizes: 4B, 12B, and 27B configurations. They support translation across 55 language pairs covering diverse language families including high-, mid-, and low-resource languages. TranslateGemma retains strong multimodal capabilities inherited from Gemma 3, enabling translation of text within images without requiring specific multimodal fine-tuning.

The models achieve exceptional efficiency through a specialized two-stage fine-tuning process. First, supervised fine-tuning uses diverse datasets including human-translated texts and synthetic translations from Gemini models. Second, reinforcement learning employs an ensemble of reward models like MetricX-QE and AutoMQM to refine translation quality for contextual accuracy and natural-sounding output.

TranslateGemma provides high-fidelity translation quality while using significantly fewer parameters than baseline models, enabling higher throughput and lower latency without sacrificing accuracy. The models are designed to serve as robust foundations for further adaptation, making them ideal starting points for researchers to fine-tune state-of-the-art models for specific language pairs.

The models are optimized for diverse deployment environments: the 4B model for mobile and edge deployment, the 12B model for consumer laptops, and the 27B model for cloud environments running on single H100 GPUs or TPUs. They are available through platforms including Kaggle, Hugging Face, Vertex AI, and the Gemma Cookbook.