Google has announced its new family of translation models, TranslateGemma, built upon the open-weighted Gemma 3 model. TranslateGemma supports 55 different languages, including widely used ones like Turkish, English, Spanish, French, Chinese, and Hindi. This broad language spectrum makes the model appealing for both developers and multilingual applications.
The timing of Google's announcement is noteworthy, as OpenAI quietly launched its ChatGPT Translate feature just a day prior. ChatGPT Translate offers an approach focused on contextual and tonal accuracy rather than word-for-word translation. However, in my tests, I found significant shortcomings, as it behaves more like a chatbot disguised as a translation tool. You can see the serious issue I mentioned in the news below.
Three Different Versions
TranslateGemma, on the other hand, is positioned more around its model architecture and performance efficiency. TranslateGemma is offered in three different versions: 4B, 12B, and 27B parameters. According to technical data shared by Google, the 12B parameter model outperforms the foundational Gemma 3 27B model in WMT24++ benchmark tests.
Google states that the 4B model is optimized for inference on mobile devices, while the 12B model is suitable for consumer-grade laptops. The largest model, the 27B version, has higher hardware requirements and is recommended to be run in a cloud environment with a powerful GPU like a single NVIDIA H100.
TranslateGemma's capabilities are not limited to text alone. Evaluations conducted in Vistra image translation benchmark tests show that the model is also successful at translating text within images. Google notes that it did not fine-tune the model specifically for this purpose.
Gemini Detail in Training
The company also explained that the high "intelligence density" in TranslateGemma is due to a two-stage special training process. In the first stage, the Supervised Fine-Tuning method was used. During this process, Gemma 3 models were trained with both texts prepared by human translators and high-quality synthetic data generated by Gemini models.
In the second stage, Reinforcement Learning was employed. In this phase, multiple reward models, including advanced measurement systems like MetricX-QE and AutoMQM, were used to make translations more natural, fluent, and contextually appropriate.
How to Use?
TranslateGemma models are available for download via the Kaggle and Hugging Face platforms for those who wish to try them or use them in their own projects. They can also be explored through the Gemma Cookbook and deployed via Vertex AI.
0 Comments: