Google releases TranslateGemma open-source translation models supporting 55 languages

Key facts

• Google released TranslateGemma, a set of open-source translation models that work across 55 languages.
• The collection comes in three sizes—4B, 12B, and 27B parameters—designed to run on different devices from smartphones to cloud servers.
• The 12B model outperforms Google's 27B baseline on the WMT24++ benchmark while using less than half the computing power.
• The 4B version matches the 12B baseline's performance, making it practical for mobile apps that need offline translation.
• Google built these models by fine-tuning Gemma 3 on a dataset that mixes human translations and synthetic text from Gemini.
• A second training phase used reinforcement learning with metrics like MetricX-QE and AutoMQM to improve how natural the translations sound.
• The models inherited the multimodal capabilities from Gemma 3 and can translate text within images without extra training.