TranslateGemma: Google releases AI model for translation

TranslateGemma is a freely available AI model family specialized in translation services. It is based on Gemma 3, Google's most powerful open model to date. TranslateGemma masters 55 different languages.

The model is available in three sizes (4B, 12B, 27B). Model distillation was used for development, meaning the knowledge of a larger model was made available for the smaller, faster model. Therefore, the medium-sized model with 12 billion parameters is expected to perform better in benchmarks focused on translation than the multimodal Gemma 3 model with 27 billion parameters, which is not specialized for translation.

Google writes in the blog post that this achieves better translation performance with lower latency, which is also more cost-effective. Despite the wide selection of training material, TranslateGemma is ideal for developers who want to build upon it with further fine-tuning and optimize translation capabilities for less frequently used languages.

Translation model with multimodal capabilities

In addition, Google writes that the models will continue to be able to follow instructions very precisely. This ability is often lost when a model is specialized. Text within images can also still be translated. This is a multimodal capability of Gemma 3.

The training of TranslateGemma was conducted in two steps: first with parallel data, i.e., human and synthetically translated texts. Subsequently, reinforcement learning was used to improve the quality of the translations. This means that evaluation models reviewed the responses, and the results were fed back to the model.

TranslateGemma is available via HuggingFace or Vertex AI, for example. Google has also published a technical report. Like its predecessors, the model does not meet the definition of open-source but can be described as open.

OpenAI introduced ChatGPT Translate just a few days ago. A closed translation service, similar to DeepL or the previous Google Translate.

(emw)