The Pulse
Google’s TranslateGemma Brings Offline Translation to 55 Languages
Google has released TranslateGemma, a three-model translation family built on Gemma 3, with evaluated results across 55 languages and on-device deployment options.
AI.info Team ·
Google is pushing translation in two directions at once: toward larger systems that handle speech, tone and context, and toward smaller models that developers can deploy across different environments. TranslateGemma, released on January 15, 2026, is the company’s new family of translation models built on Gemma 3 and designed to work across 55 languages.
The release emphasizes a tradeoff between model size and deployment target. Google positions the models for mobile and edge devices, consumer laptops and cloud-based workloads, while presenting their efficiency as a way to bring stronger translation performance to smaller systems.
“Today, we're introducing TranslateGemma, a new collection of open translation models built on Gemma 3, helping people communicate across 55 languages, no matter where they are or what device they own.”
David Vilar, Staff Research Scientist, and Kat Black, Product Manager, Google
Three models, three deployment targets
TranslateGemma comes in 4-billion-, 12-billion- and 27-billion-parameter versions. Google positions the 4B model for mobile and edge deployment, the 12B model for consumer laptops and the 27B model for higher-fidelity workloads that can run on a single H100 GPU or TPU in the cloud.
All three models are built on Gemma 3. TranslateGemma retains Gemma 3’s multimodal capabilities, allowing it to translate text within images as well as handle ordinary text translation. Developers can download the models through Hugging Face and Kaggle, experiment through the Gemma Cookbook, or deploy them in Vertex AI.
Google claims a smaller model beats its larger baseline
Google’s main performance claim concerns the 12B model. On the company’s evaluation using the WMT24++ benchmark and MetricX, TranslateGemma 12B outperformed the Gemma 3 27B baseline while using less than half as many parameters.
The WMT24++ evaluation covers 55 languages from high-, mid- and low-resource language groups. Google says TranslateGemma reduced the error rate against the baseline Gemma model in every language tested. The company also says its 4B model rivals the performance of the larger 12B baseline, supporting its case for mobile inference.
Google does not present the benchmark as proof that every translation task will perform equally well. The company says it trained and evaluated the models on 55 language pairs, while nearly 500 additional pairs received training but do not yet have confirmed evaluation metrics. That distinction matters for developers working with languages outside the tested set.
Gemini data feeds an open model
TranslateGemma’s training process combines human-translated material with synthetic translations generated by Gemini models. Google describes the process as a two-stage program: supervised fine-tuning on parallel data, followed by reinforcement learning guided by reward models including MetricX-QE and AutoMQM.
Google says the training data was designed to improve coverage and fidelity for low-resource languages, where the supply of high-quality parallel text is often limited. The company also says the models’ image-translation gains came without specific multimodal fine-tuning during TranslateGemma training.
The models are available for researchers and developers to download and adapt, rather than only through a hosted translation service. Their distribution is subject to Google’s Gemma Terms of Use and associated use restrictions; it is not an unrestricted open-source license.
Offline access meets a hardware barrier
Google’s September 15, 2026 post, “AI for everyone in every language,” presents TranslateGemma as part of an effort to serve people with limited or intermittent connectivity. Google says more than 3 billion people lack reliable internet access and describes the models as running efficiently on-device, without requiring a cloud or internet connection for high-quality translation.
The same post says that powerful AI models still require capable hardware, excluding hundreds of millions of people who use feature phones in low-resource regions. Google says it supports organizations including Viamo, whose Ask Viamo Anything voice assistant brings Gemini capabilities to standard feature phones. Google says Viamo piloted the service in Rwanda and that it had used Gemini to answer more than 2 million questions.
TranslateGemma therefore addresses dependence on a live connection while leaving hardware access as a separate constraint. Its three model sizes give developers options for different deployment environments, but Google’s measured performance claims apply to the 55 language pairs it evaluated.
What developers can use now
TranslateGemma is available through Google’s published download and deployment channels rather than as a new consumer mode inside Google Translate. Developers can inspect the models on Hugging Face or Kaggle, adapt them for specific language pairs, test image translation, and deploy the larger version through Vertex AI.
Google also reports training on nearly 500 extra language pairs beyond the 55 it evaluated. Those pairs are presented as a starting point for further research, not as a performance guarantee. For now, the concrete release is a three-model family with measured results across 55 language pairs and a 4B version aimed at mobile and edge deployment.