What is TranslateGemma?

TranslateGemma is a new suite of open AI translation models built on Google’s Gemma 3. Available in 4B, 12B, and 27B parameter sizes, it enables high-quality communication across 55 languages while maintaining exceptional efficiency. These models are designed to run on mobile devices, local hardware, and cloud environments without sacrificing translation quality. You can download TranslateGemma on Kaggle or Hugging Face and deploy it in Vertex AI.

Who it's for

• Mobile developers who need lightweight, high-performance translation models that run efficiently on-device without cloud dependency.
• Enterprise teams deploying multilingual customer support or content localization at scale, seeking open models with strong accuracy and low latency.
• AI researchers and hobbyists exploring open translation technology who want to fine-tune or integrate state-of-the-art models into custom workflows.

Key features

Outperforms models twice its size

The 12B TranslateGemma model outperforms the Gemma 3 27B baseline on the WMT24++ benchmark, using less than half the parameters. This efficiency breakthrough means higher throughput and lower latency without sacrificing accuracy. The 4B model rivals the performance of the larger 12B baseline, making it a powerful choice for mobile inference.

Built from Gemini distillation

TranslateGemma uses a specialized two-stage fine-tuning process. First, supervised fine-tuning on a diverse dataset of human-translated and high-quality synthetic translations from Gemini models ensures broad language coverage. Then, reinforcement learning with an ensemble of reward models (including MetricX-QE and AutoMQM) refines translations to be more contextually accurate and natural-sounding.

Unprecedented language coverage

TranslateGemma is rigorously trained and evaluated on 55 language pairs, covering major languages like Spanish, French, Chinese, and Hindi, as well as many low-resource languages. This broad support ensures reliable performance across diverse linguistic contexts.

Retains multimodal capabilities

TranslateGemma keeps the strong multimodal abilities of Gemma 3, enabling translation of text within images. This makes it suitable for applications like document scanning, signage translation, and visual content localization.

What stands out

"The 12B TranslateGemma model outperforms the Gemma 3 27B baseline using less than half the parameters."

This efficiency is the product’s defining edge. By distilling knowledge from Google’s most advanced large models into compact, open architectures, TranslateGemma delivers high-fidelity translation quality that rivals much larger systems. Developers get a model that runs faster, uses fewer resources, and still produces accurate, natural-sounding translations across 55 languages.

Worth checking out if…

You need an open, efficient translation model that runs on mobile or local devices without cloud dependency. TranslateGemma is ideal if you value parameter efficiency and want to deploy high-quality multilingual support at scale, whether for apps, websites, or research projects.