Google has released a new suite of open translation models based on its open-weight Gemma 3 model. Google is calling the suite a "significant step forward in open translation". The new collection, TranslateGemma, can handle up to 55 languages, including the popular ones you"d expect, like Spanish, French, Chinese, and Hindi.
This announcement came just hours after OpenAI released ChatGPT Translate. ChatGPT Translate focuses more on translating with the correct tone and context, rather than just doing a direct word-for-word swap. It features a two-panel interface with auto-language detection on the left and the target language on the right, kind of similar to tools like Google Translate.
Anyway, TranslateGemma is available in 4B, 12B, and 27B parameter sizes. According to Google, the 12B model outperforms the baseline Gemma 3 27B model when measured on the WMT24++ benchmark. For developers, getting that kind of performance from a model with less than half the parameters means higher throughput and lower latency without sacrificing accuracy.
Google said the 4B model is optimized for mobile inference and that the 12B model is fit for consumer laptops. The big 27B model, however, needs more power, like a single NVIDIA H100 in the cloud. Tests on the Vistra image translation benchmark showed TranslateGemma could better translate text within images, even though it was not specifically fine-tuned for it.
Google also noted how it achieved this density of intelligence through a specialized two-stage process. First, it used Supervised Fine-Tuning, training the base Gemma 3 models on a mix of human-translated texts and high-quality synthetic data generated by Gemini models.
After that, it implemented a Reinforcement Learning phase. This second stage used a group of reward models, including advanced metrics like MetricX-QE and AutoMQM, to guide the models toward more natural and contextually accurate translations.
The TranslateGemma models are available for download on Kaggle and Hugging Face for anyone who wants to experiment with them or build on top of them.