German consortium releases Soofi S, an open 30B Mamba-MoE model it says tops open rivals in two languages

A German public-private consortium has released Soofi S, a fully open 30-billion-parameter language model its makers say tops other open models on both German and English benchmarks. The SOOFI Consortium (Sovereign Open Source Foundation Models) published Soofi-S-30B-A3B on July 13, 2026, according to its Hugging Face model card, and it was still drawing wide developer discussion three days later.

The release is a sovereignty play. It is funded by the German Federal Ministry for Economic Affairs and Energy under the EU’s IPCEI-CIS program and coordinated by the industry group KI Bundesverband, a bid for European control over foundation models rather than dependence on U.S. labs.

The architecture is the technical draw. Soofi S is a hybrid Mixture-of-Experts model that pairs Mamba-2 state-space layers with a smaller set of attention layers — 128 routing experts, six activated per token — for about 30 billion total parameters and roughly 3.5 billion active during inference. Named members include the Fraunhofer institutes IAIS and IIS, the German Research Center for Artificial Intelligence (DFKI) and TU Darmstadt. Training ran across about 27 trillion tokens and consumed roughly 253,000 GPU-hours on up to 512 Nvidia B200 chips on Deutsche Telekom’s Industrial AI Cloud in Munich.

The consortium’s pretraining report claims Soofi S leads all fully open models, ahead of OLMo 3 32B and Apertus 70B, with 73.8% on the HumanEval coding benchmark and 84.2% on a German MBPP variant. Those are self-reported figures that have not been independently verified.

One caveat sits on the download page itself: the model card labels this an unaligned base checkpoint with no instruction tuning or safety alignment. The scores describe raw pretraining capability, not a model anyone should deploy as an assistant without further work.

Whether an aligned, instruction-tuned version follows will determine if Soofi S becomes a usable European alternative or stays a benchmark demonstration.