Google Introduces TranslateGemma Open Translation Models

Google has introduced TranslateGemma, a new collection of open translation models built on Gemma 3. The models are designed to enable high-quality translation across 55 languages and are available in three sizes: 4 billion, 12 billion, and 27 billion parameters.

TranslateGemma models were developed using a two-stage fine-tuning process that distills the capabilities of Gemini models into a compact architecture. The process includes supervised fine-tuning on a mix of human and synthetic parallel data, followed by reinforcement learning guided by multiple reward models. This approach improves contextual accuracy and naturalness in translations.

In testing, the 12B TranslateGemma model outperformed the Gemma 3 27B baseline on the WMT24++ benchmark, showing higher translation quality with fewer parameters. The 4B model delivers comparable performance to the larger 12B baseline, making it suitable for mobile and edge deployment. All models demonstrated reduced error rates across high-, mid-, and low-resource languages.

TranslateGemma also retains multimodal capabilities from Gemma 3, improving performance on image translation benchmarks without specific multimodal training. The models are optimized for different environments: the 4B model for mobile devices, 12B for laptops, and 27B for cloud-based GPUs or TPUs. Researchers and developers can access TranslateGemma through Kaggle, Hugging Face, Vertex AI, and the Gemma Cookbook.

The release provides an open foundation for further adaptation and fine-tuning, allowing the community to build translation systems tailored to specific language pairs or specialized use cases.