The tech giant Google has introduced TranslateGemma, a new collection of open translation models built on Gemma 3, available in 4B, 12B, and 27B parameter sizes.
It represents a significant step forward in open translation, helping people communicate across 55 languages, no matter where they are or what device they own.
By distilling the knowledge of Google’s most advanced large models into compact, high-performance open models, it has created a suite where efficiency doesn’t require a compromise on quality.
The most remarkable finding in the technical evaluation is the efficiency of these models. Through Google’s specialized training process, the 12B TranslateGemma model outperforms the Gemma 3 27B baseline as measured using MetricX on the WMT24++ benchmark.
For developers, this is a massive win. They can achieve high-fidelity translation quality using less than half the parameters of the baseline model. This efficiency breakthrough allows for higher throughput and lower latency without sacrificing accuracy. Similarly, the 4B model rivals the performance of the larger 12B baseline, making it a powerful model for mobile inference.
The TranslateGemma model has been trained on nearly 500 additional language pairs. It is designed for TranslateGemma to serve as a robust foundation for further adaptation, making it an ideal starting point for researchers to fine-tune their own state-of-the-art models for specific language pairs or to improve quality for low-resource languages.
TranslateGemma models are available to download on Google’s Hugging Face listing and Kaggle’s website. Additionally, developers and enterprises can also access them via Vertex AI, the company’s cloud-based AI hub. These models are available with a permissive license allowing both academic and commercial use cases.
