Google has released TranslateGemma, a set of open translation models based on its open-weighted Gemma 3 model. The company describes the release as “a significant step forward in open translation.” The new family of models supports 55 languages, including Ukrainian.
Google has released TranslateGemma, a set of open translation models based on its open-weighted Gemma 3 model. The company describes the release as “a significant step forward in open translation.” The new family of models supports 55 languages, including Ukrainian.
Just yesterday we wrote about the launch of the AI translation feature in ChatGPT , and now OpenAI's main competitor — Google — has introduced its new product.
TranslateGemma is available in 4, 12, and 27 billion parameter versions. According to Google, the 12B model outperforms the base Gemma 3 27B model on the WMT24++ benchmark. For developers, this performance of a model with half the parameters means higher throughput and lower latency without sacrificing accuracy.
Google says the 4B model is optimized for mobile devices, while the 12B version is ideal for consumer laptops. The powerful 27B model, on the other hand, requires more serious resources, such as a single NVIDIA H100 GPU in the cloud.
Tests on the Vistra image translation benchmark showed that TranslateGemma is better at translating text in pictures, even though it was not trained specifically for this task, Neowin writes .
Google explained that this performance was achieved through a specialized two-stage process. The first stage involved supervised fine-tuning: the base Gemma 3 models were trained on a mixture of human-translated texts and high-quality synthetic data generated by Gemini models.
This was followed by a reinforcement learning phase. This second phase used a set of reward models, including advanced metrics such as MetricX-QE and AutoMQM, to guide the models towards more natural and contextually accurate translations.
TranslateGemma models are available for download on Kaggle and Hugging Face.
