Gemma 4 is a huge improvement in many European languages, including Danish, Dutch, French and Italian

Google's latest iteration of its open-source Gemma models, Gemma 4, is making waves for its surprisingly strong performance across multiple European languages. According to new benchmark data from EuroEval, the Gemma 4 31B parameter model is competing with—and often beating—much larger proprietary models. It ranks first in Finnish, second in Danish, French, and Italian, and places within the top three for Dutch, English, and Swedish. This performance is notable because Gemma 4 models are relatively small and efficient compared to the multi-trillion parameter behemoths they are outperforming in these specific linguistic tasks.

The benchmarks suggest Google has made significant strides in multilingual training and data curation for this release. The strong showing across Germanic (Dutch, Danish, Swedish, English), Romance (French, Italian), and Uralic (Finnish) language families indicates broad linguistic capability. For developers and companies in Europe, this means a powerful, open-source alternative is now available for building AI applications that require nuanced understanding of local languages, potentially at a lower cost and with greater customization potential than closed APIs from OpenAI or Anthropic.

While benchmarks provide a standardized measure, the real-world test will be how these models perform in production applications like customer service chatbots, document translation, and content generation. The open-source nature of Gemma 4 allows for fine-tuning on specific domains or regional dialects, a key advantage for European businesses. If the performance holds up, it could accelerate AI adoption across the continent by providing a high-quality model that respects data sovereignty and reduces dependency on US-based AI infrastructure.