TranslateGemma
2026-01-15 · Publisher : Google · Domain : Artificial Intelligence
Google has released TranslateGemma, a family of open translation models in 4B, 12B and 27B parameters built on Gemma 3, evaluated on 55 languages and designed to run on a laptop or mobile device.
What changes
- The 12B model outperforms the Gemma 3 27B base according to MetricX on WMT24++, with less than half the parameters.
- The 4B rivals the 12B base, which makes it suitable for mobile inference.
- Two-stage training: supervised fine-tuning on parallel data (human and synthetic Gemini data) followed by reinforcement learning (MetricX-QE, AutoMQM).
- Rigorous evaluation on 55 language pairs, plus close to 500 additional pairs trained without confirmed metrics.
- Gemma 3's multimodal capabilities are retained (translation of text within images tested on Vistra).
### Specifications - Sizes: 4B, 12B and 27B parameters. - Modalities: text and image as input, text as output. - Availability: open weights (Hugging Face, Kaggle), local or cloud execution. - Context and knowledge cutoff: not disclosed in the blog post. ### Pricing Open weights, no API pricing disclosed. ### Benchmarks Scores published by the vendor: the blog post gives relative comparisons on WMT24++ (MetricX, lower is better) with no absolute figure in the text. No score table included here. ### Key takeaways An open, lightweight option for embedded or internally hosted machine translation, without depending on a third-party API.
Links
Tags : Open Source, Google, Gemma, TranslateGemma, Traduction