German consortium releases Soofi S, a 31.6B open MoE model

The interesting bit in this week's release from a German research consortium isn't the leaderboard number, it's the assembly. The Decoder reports that Soofi S 30B-A3B, coordinated by the KI Bundesverband and built by a group that includes the Fraunhofer Institutes IAIS and IIS, DFKI, TU Darmstadt, the University of Würzburg, and AI companies Ellamind and Merantix Momentum, was trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich. Between March and May the team ran up to 512 Nvidia B200 GPUs for about 253,000 GPU-hours, processing roughly 27 trillion tokens across three phases.

Architecturally, the consortium didn't try to reinvent anything. They adopted Nvidia's Nemotron 3 Nano design without modification, a hybrid mixture-of-experts that pairs Mamba-2 layers with standard attention and activates only 3.2 of its 31.6 billion parameters per generated token. Michael Fromm, described as part of the project's technical leadership, and colleagues report aggregate scores of 79.1 on German benchmarks and 70.1 on English, ahead of OLMo 3 32B from the Allen Institute for AI and Apertus 70B from ETH Zurich and EPFL among fully open peers, with 73.8 percent on HumanEval and 70.2 on MBPP.

The strategic angle is the openness posture. The team is releasing model weights along with selected intermediate checkpoints and the complete training and evaluation code, and claims Soofi S meets the Open Source AI Definition 1.0 from the Open Source Initiative, a bar higher than the 'open weights' label most labs stop at. For European buyers who have spent two years hearing there is no serious sovereign option, that combination of a domestic training stack and a genuinely reconstructible model is more interesting than any single benchmark line.

The honest caveats sit in the same report. Beyond 32,000 tokens of context Soofi S's hit rate on RULER-style extraction drops to around 3 percent, math trails Qwen3.5 35B-A3B on Minerva MATH-DE (56 versus 76.5), and the exact license hasn't been finalized yet. What the reporting doesn't give you is the inference cost per token on commodity hardware, or whether a larger Soofi M or L is funded. Still, take the specifics as reported, not settled, and the direction is the part worth watching.