Google has unveiled TranslateGemma, a new family of open-source machine translation models built on top of Gemma 3. Designed to support communication across 55 languages, the models aim to provide high-quality translation while remaining efficient and accessible to developers and researchers worldwide.
The TranslateGemma family is released in three sizes: 4B, 12B, and 27B parameters. The smallest model targets mobile and edge devices, while the 12B model is optimized for consumer laptops and local hardware. The largest 27B model is intended for production-scale deployments on GPUs and TPUs, offering flexibility across different use cases.
According to Google’s technical documentation, the models were trained using a two-stage approach. First, supervised fine-tuning was applied using a mix of high-quality human translations and synthetic data generated by Gemini models. This was followed by a reinforcement learning phase focused on producing more natural and fluent translations, rather than rigid, robotic outputs.
Beyond text-to-text translation, TranslateGemma retains multimodal capabilities, enabling text extraction and translation directly from images. Its ability to run fully offline and keep data on-device positions it as a strong alternative to cloud-based translation services, particularly for privacy-sensitive applications, travel scenarios, and enterprise environments.