Google has officially introduced TranslateGemma, a new suite of open-weight translation models built upon its advanced Gemma 3 architecture. This significant launch aims to deliver high-quality and efficient machine translation across 55 languages, encompassing both major global languages like Spanish, French, and Chinese, as well as several low-resource languages, including Indian vernaculars. TranslateGemma is available in three parameter sizes: 4B, 12B, and 27B, allowing for flexible deployment across various devices, from smartphones and consumer laptops to high-end cloud infrastructure. A notable achievement highlighted by Google is that the 12B model can outperform the larger Gemma 3 27B baseline model on the WMT24++ benchmark, while requiring significantly fewer computational resources. This efficiency makes advanced translation more accessible for local and self-hosted applications. The timing of TranslateGemma's release is particularly strategic, coming just hours after OpenAI debuted its dedicated ChatGPT Translate tool. This parallel launch by two tech giants clearly intensifies the rivalry in the rapidly evolving AI translation landscape, marking a new phase where both companies are pushing distinct approaches to AI-powered language services. While ChatGPT Translate emphasizes flexibility, tone control, and conversational rewriting across over 50 languages, Google's TranslateGemma focuses on open, efficient models designed for broad deployment and developer control. The development of TranslateGemma involved a specialized two-stage fine-tuning process. This process includes Supervised Fine-Tuning on a diverse dataset of human-translated texts and high-quality synthetic data generated by Gemini models, followed by a Reinforcement Learning phase to further refine translation quality using advanced metrics. Furthermore, TranslateGemma models retain the multimodal capabilities inherited from Gemma 3, allowing them to translate text embedded within images, even without specific multimodal fine-tuning for this feature during their development. The models are readily accessible to developers and researchers through platforms like Hugging Face, Kaggle, the Gemma Cookbook, and Vertex AI.