Sovereign AI development is gaining relevance across Europe as industry and organizations seek greater control over model training, deployment, and data governance. A new contribution from a consortium of German research institutions focuses on combining openness with computational efficiency. All technical details about the model training are published in a brand-new Tech Report.

Abstract

The consortium presents Soofi S 30B-A3B, an open, sovereign AI language model from Germany. The consortium consists of German research institutions, who trained the model entirely on German HPC infrastructure. It combines a hybrid Mamba-Transformer architecture with a Mixture-of-Experts approach, activating only around 3 of 30 billion parameters per input, enabling strong performance at low inference cost.

Trained on 27 trillion text tokens in German and English, Soofi S matches significantly larger international models on independent benchmarks and outperforms all previous European sovereign language models. The consortium released the model weights, full training documentation, and code under open licenses.

Full Technical Report