A German research consortium, coordinated by the KI Bundesverband (German AI Association), has released Soofi S 30B-A3B, an open language model. This model reportedly achieves top scores on English and German benchmarks among fully open models, surpassing predecessors like OLMo 3 32B and Apertus 70B. Soofi S utilizes a hybrid Mamba-Transformer architecture, allowing it to maintain high throughput even with very long contexts, unlike dense models that experience significant performance drops.

With 31.6 billion total parameters, Soofi S activates only about 3.2 billion per token, making its compute cost comparable to a 3B model. The model was trained on approximately 27 trillion tokens on Deutsche Telekom's Industrial AI Cloud in Munich. The training mix was deliberately weighted toward German, with German data comprising a significantly larger share than in comparable models. This focus on German data improved language proficiency by over 15 points compared to the Nemotron baseline.

While Soofi S excels in many benchmarks, including code generation and German-specific knowledge, it shows weaknesses in German competition math and open factual retrieval. The consortium, funded by the German Federal Ministry for Economic Affairs and Energy, aims to develop an open European AI model family for sovereign infrastructure. Michael Fromm, part of the project's technical leadership, argues that traditional scaling laws do not apply to Mixture-of-Experts (MoE) architectures, citing Nvidia's training data volume as a comparison.