TranslateGemma comprises three model sizes—4B, 12B, and 27B parameters—engineered to offer high-fidelity translation across diverse language pairs including low-resource languages. The 12B model notably outperforms the 27B Gemma 3 baseline while using less than half the parameters, indicating significant efficiency improvements without sacrificing translation quality.
Trained through specialized supervised fine-tuning and reinforcement learning leveraging advanced metrics, TranslateGemma extends broad language coverage and multimodal capabilities, such as translating text within images. While evaluation confirmed gains on the WMT24++ benchmark, extended language support remains under exploration pending further performance validation.
Source & further reading
TranslateGemma: A new suite of open translation models ↗Source published 2026-01-15. This report was prepared with AI assistance from the linked primary-source material. Claims and limitations should be checked against the original announcement or research.