German Research Consortium Releases Open LLM 'Soofi S'

A German research consortium has released the large language model 'Soofi S 30B-A3B'. The model was developed from scratch on Deutsche Telekom's cloud infrastructure located in Munich, Germany. It has been released as a fully open model and achieves top-tier performance among comparable open models in both English and German benchmarks.

In the world of large language models, models trained primarily on English overwhelmingly dominate. Models that demonstrate high performance on languages other than English, particularly European languages like German, remain scarce, and multilingual support has long been a recognized challenge. Additionally, from the perspective of geographically distributed development including cloud infrastructure, the reality is that major AI research hubs are concentrated in the United States and China. This effort stands apart from such trends and positions itself within the context of Europe-led AI development.

Soofi S has a total parameter count of 33.6 billion, but employs a hybrid architecture where only a portion of parameters are actually used in each processing step (per token). This mechanism, also called 'sparse activation', selectively computes only necessary parts rather than fully engaging all parameters each time. Consequently, it exhibits the characteristic of maintaining processing speed even as text length increases. The training data deliberately emphasizes German content, positioning German language support as the centerpiece of the development design.

In benchmarks—standard evaluation tests for comparing model performance—the model achieved results surpassing fully open competing models in both English and German. The significance of these results is further amplified by the fact that, as a fully open model, researchers and developers can freely inspect and improve the model's internals.

The significance of this model extends beyond technical performance. The fact that European research institutions have independently developed a competitive model using regional infrastructure without relying on major U.S. cloud providers is noteworthy in the geopolitical context of AI. At a time when the EU is advancing AI regulation, the demonstration of strong performance in major languages by a European open model represents an important step toward securing autonomous AI development capabilities. This achievement carries substantial weight from the perspective of reducing concentrated risk in AI development power.

A key question going forward is how this model will be utilized in actual operations and research. The emergence of an open model with strong German capabilities has the potential to broaden the base for German-language AI applications in specialized domains such as administration, healthcare, and law. Additionally, as a demonstration of the practical utility of sparse activation architecture, it holds reference value for other research institutions and developers. It will be worth continuing to observe how Europe's AI development community receives this achievement and carries it forward into subsequent efforts.