|
Google Launches TranslateGemma Open Translation Models Based on Gemma 3
Google announces the launch of TranslateGemma on January 15, 2026, as an open suite of translation-focused AI models derived from Gemma 3. The models provide efficient, high-quality translation across 55 languages and are designed to run across devices from mobiles to cloud systems.
TranslateGemma (Open Translation Models) :
| Dimension | Key Details |
|---|---|
| Type | Provides for a suite of open, translation-focused artificial intelligence models. |
| Base architecture | Built on the Gemma 3 architecture. |
| Language coverage | 55 languages. |
| Model sizes | Available in 4B, 12B and 27B parameter sizes. |
| Cross-device deployment | Applies to translation across devices from mobiles to cloud systems. |
| Training methodology | Uses a two-stage fine-tuning process with supervised learning on human and synthetic translations, followed by reinforcement learning guided by multiple reward models. |
| Benchmark evaluation | Technical evaluations use the WMT24++ benchmark. |
| Benchmark performance claims | On WMT24++, the 12B TranslateGemma model outperforms the 27B Gemma 3 baseline; the 4B model rivals the earlier 12B baseline. |
| Efficiency attributes | Provides for faster inference, lower latency, and reduced computational cost while maintaining high translation fidelity. |
| Language-pair expansion | Training has been extended to nearly 500 more language pairs. |
| Multimodal capability | Retains multimodal strengths of Gemma 3 and shows improved performance in translating text within images even without dedicated multimodal fine-tuning. |
| Suggested deployment by size | 4B applies to mobile and edge devices; 12B applies to consumer laptops and local development; 27B applies to high-fidelity cloud deployment on GPUs or TPUs. |