German Consortium Debuts Soofi S, an Open 30B MoE Model

A German research consortium has released Soofi S Base, an openly available large language model built on a mixture-of-experts design with roughly 30 billion parameters. The model is available on Hugging Face, and according to The Decoder it posts leading scores on both English and German evaluations.

What sets Soofi S apart is its architecture. Rather than the standard transformer, it leans on Mamba-2, a state-space approach that promises more efficient handling of long sequences. Pairing that with a mixture-of-experts layout means only a fraction of the model's parameters activate for any given token, keeping inference costs lower than the headline count suggests.

Why it matters

Europe has been vocal about wanting sovereign, openly licensed models trained with local languages in mind, and Soofi S is a concrete step in that direction. A model that competes in German without sacrificing English performance is a meaningful signal for developers and institutions that need strong multilingual behavior.

• Roughly 30B parameters in a sparse MoE configuration
• Built on the Mamba-2 state-space architecture
• Reported strength across English and German benchmarks
• Released openly for the community to test and build on

As always, independent verification will matter. Benchmark leadership claims are worth watching once the wider community runs its own evaluations, but the combination of an alternative architecture and a bilingual focus makes Soofi S a release worth tracking.