TranslateGemma is a collection of open translation models built on Gemma 3, available in 4B, 12B, and 27B parameter sizes. It represents a significant advancement in open translation technology, designed to help people communicate across 55 languages efficiently without compromising quality.
Key Features:
- Efficient Performance: The 12B model outperforms the Gemma 3 27B baseline using less than half the parameters, enabling higher throughput and lower latency.
- Multilingual Support: Trained and evaluated on 55 language pairs, including major and low-resource languages, with extended training on nearly 500 additional pairs for community adaptation.
- Multimodal Capabilities: Retains Gemma 3's ability to translate text within images, as demonstrated on the Vistra image translation benchmark.
- Flexible Deployment: Available in three sizes optimized for different environments: 4B for mobile/edge, 12B for consumer laptops, and 27B for cloud deployment on single H100 GPU or TPU.
- Open Access: Downloadable on Kaggle and Hugging Face, deployable in Vertex AI, with a technical report available for research.
Use Cases:
- Developers and researchers building translation applications or adapting models for specific language pairs.
- Mobile and edge deployment for on-device translation without internet connectivity.
- Cloud-based translation services requiring high fidelity and efficiency.
- Multimodal applications needing image text translation alongside standard text translation.
