Google Unveils ‘TranslateGemma’ Open Translation Models
A leading tech company has launched TranslateGemma, a new set of open translation models built on Gemma 3. The collection comes in 4B, 12B, and 27B parameter sizes.
Notably, the models support communication across 55 languages on a wide range of devices. As a result, users can translate more effectively regardless of location or hardware.
Additionally, the models deliver strong performance while staying compact. Instead of relying only on larger systems, the team distilled knowledge from advanced large models into smaller open models. Therefore, the suite aims to boost efficiency without compromising translation quality.
Strong results despite fewer parameters
The technical evaluation highlights how efficient the models are. Specifically, the 12B TranslateGemma model outperforms the Gemma 3 27B baseline when measured with MetricX on the WMT24++ benchmark.
For developers, this is a major advantage. For instance, they can achieve high-fidelity translations with less than half the parameters of the baseline model. Consequently, the models can provide higher throughput and lower latency while maintaining accuracy.
Meanwhile, the 4B model performs competitively against the larger 12B baseline. Because of this, it becomes a practical option for mobile inference and other lightweight deployments.
Built for adaptation and broad access
The TranslateGemma model has been trained on nearly 500 additional language pairs. Moreover, it is designed to serve as a strong foundation for further adaptation.
As a result, researchers can fine-tune models for specific language pairs or improve translation quality for low-resource languages. In addition, the open nature of the release supports experimentation across many use cases.
TranslateGemma models are available to download on the Hugging Face listing and Kaggle’s website. Furthermore, developers and enterprises can access them through a cloud-based AI hub. Finally, the models come with a permissive license that supports both academic and commercial use cases.
