Google's TranslateGemma open AI models can translate text, images: Details

Google introduces TranslateGemma, a collection of open translation models offering 55-language support through Gemma 3-based models

Google has introduced TranslateGemma, a new collection of open translation AI models built on Gemma 3. The company said that the models support translation across 55 languages and are available in three sizes: 4B, 12B, and 27B parameters. According to Google, TranslateGemma is part of its efforts to advance open translation models by distilling knowledge from its Gemini models into smaller, open architectures.

What is TranslateGemma

TranslateGemma is a family of AI models focused specifically on language translation. Google said that the models were created using a specialised training process that transfers capabilities from its Gemini models into the Gemma 3 framework.

 

 

The company has released TranslateGemma in three sizes to support different deployment needs, ranging from smaller environments to cloud-based systems.

Language coverage and translation quality

Google said TranslateGemma has been trained and evaluated across 55 languages, including widely spoken languages such as Spanish, French, Chinese, and Hindi, as well as several low-resource languages.

 

In its internal evaluations, Google said that TranslateGemma reduced translation errors compared to baseline Gemma models across all tested language groups. The company also trained the models on nearly 500 additional language pairs to allow further research and adaptation, though evaluation results for those pairs have not yet been confirmed.

 

Image text translation

TranslateGemma retains the multimodal capabilities of Gemma 3. Google said testing showed that improvements in text translation also carried over to translating text within images, even without additional multimodal training during the TranslateGemma process.

Where TranslateGemma can run

Google said TranslateGemma has been designed for different environments depending on model size:

• 4B model: Intended for mobile and edge deployment
• 12B model: Designed to run on consumer laptops
• 27B model: Built for cloud deployment, capable of running on a single GPU or TPU

Availability

TranslateGemma is available through Kaggle, Hugging Face, and Vertex AI, along with technical documentation provided by Google. 

Why it matters

TranslateGemma is aimed primarily at researchers and developers working on translation systems. While end users will not interact with the models directly, Google said that they can serve as a foundation for building translation tools that support a wide range of languages across different devices.

OpenAI introduces ChatGPT Translate

In related news, OpenAI has introduced ChatGPT Translate, a dedicated translation tool that supports conversions between over 50 languages. The company said that the service can handle translations for text, spoken input, and images, while preserving context, tone, and commonly used expressions.

 

The layout closely resembles traditional translation tools, featuring side-by-side input and output fields with automatic language detection. In addition to basic translation, users can continue the conversation by asking for clarifications, requesting alternative phrasing, or changing the target language without starting a new session.