Google releases TranslateGemma open-source translation models

Google introduced TranslateGemma, a collection of open-source translation models based on Gemma 3 architecture, available in 4B, 12B, and 27B parameter configurations. The models support translation across 55 languages and are designed for deployment on various devices and platforms.

The company reports that the 12B TranslateGemma model outperforms the larger Gemma 3 27B baseline model when measured using MetricX on the WMT24++ benchmark. The 4B model demonstrates performance comparable to the 12B baseline model, according to Google's testing.

Google developed the models through a two-stage fine-tuning process. The first stage involved supervised fine-tuning on parallel translation data, including human-translated texts and synthetic translations generated by Gemini models. The second stage used reinforcement learning with reward models including MetricX-QE and AutoMQM to refine translation quality.

Testing on the WMT24++ dataset showed reduced error rates across all 55 languages compared to baseline Gemma models. The dataset covers high-, mid- and low-resource languages from various language families.

The models retain multimodal capabilities from Gemma 3, with testing on the Vistra image translation benchmark showing improved performance in translating text within images without specific multimodal training.

Google trained the models on nearly 500 additional language pairs beyond the core 55, though evaluation metrics for the extended set are not yet confirmed. The models are available for download through Kaggle, Hugging Face, and deployment via Vertex AI.

The 4B model targets mobile and edge deployment, the 12B model is designed for consumer laptops, and the 27B model is built for cloud deployment on single H100 GPU or TPU systems.

You May Also Be Interested In