German AI Alliance Releases Open-Source Soofi S 30B, Leading in EN-DE Benchmarks

Soofi S 30B-A3B was developed by a German research alliance and fully trained on German Telekom cloud infrastructure in Munich. It uses an efficient hybrid architecture that activates only a fraction of its 31.6 billion parameters per token, maintaining stable throughput even with extremely long contexts. The training data is deliberately biased towards German, enabling the model to outperform all fully open competitors on both German and English benchmarks. For users needing open-source models for German text or long contexts, Soofi S provides a balance of performance and efficiency. Its hybrid architecture reduces inference costs, while the German-heavy training data suggests superior performance on German NLP tasks compared to similar open-source models.