Google Unveils TranslateGemma: Open-Source Translation Models for 55 Languages
Google has launched TranslateGemma, a new set of open-source translation models. These models work across 55 different languages. The collection comes in three distinct sizes to suit various devices.
Three Model Sizes for Different Devices
The models are available with 4 billion, 12 billion, and 27 billion parameters. Google designed them to run on everything from smartphones to cloud servers. This flexibility allows developers to choose the right model for their needs.
The 12 billion parameter model delivers impressive performance. It outperforms Google's own 27 billion parameter baseline on the WMT24++ benchmark. Remarkably, it achieves this while using less than half the computing power.
For developers, this efficiency is a game-changer. You can now run high-quality translation locally on a standard laptop. This eliminates the need to route all translation requests through cloud APIs, which can save time and resources.
The smallest 4 billion parameter model matches the performance of the 12 billion baseline. This makes it highly practical for mobile applications that require reliable offline translation capabilities.
Training with Human and AI-Generated Data
Google built these models by fine-tuning its existing Gemma 3 framework. The training process combined human translations with synthetic text generated by the Gemini AI system.
A second phase of training employed reinforcement learning. Google used metrics like MetricX-QE and AutoMQM to refine the models. This approach helped improve how natural and fluent the translations sound to human readers.
Broad Language Coverage and Improved Accuracy
The models cover major global languages including Spanish, French, and Mandarin. They also include several lower-resource language options, expanding access to translation technology.
Google trained the models on nearly 500 additional language pairs beyond the core 55. However, the company has not formally tested these extra pairs yet. Initial testing shows reduced error rates across all 55 core languages compared to the base Gemma models.
Multimodal Capabilities for Image Text Translation
TranslateGemma inherits multimodal capabilities from the Gemma 3 architecture. Testing on the Vistra benchmark demonstrates that the models can translate text found within images.
This feature proves useful for translating signs, menus, or photographs of documents. Interestingly, Google did not specifically train the models for this image translation capability during development. The ability emerged naturally from the underlying architecture.
Availability and Deployment Options
The models are now available on popular platforms including Kaggle and Hugging Face. Each size serves different deployment scenarios:
- The 4 billion parameter version runs efficiently on mobile devices
- The 12 billion parameter model fits comfortably on consumer laptops
- The 27 billion parameter model requires a single H100 GPU for cloud deployment
This release represents a significant step in making advanced translation technology more accessible. Developers can now implement sophisticated multilingual capabilities without relying exclusively on cloud services.
