German consortium releases Soofi S, an open 30B hybrid Mamba-Transformer model

Sources: Soofi S base model (Hugging Face), pretraining report, The Decoder, HN discussion
Summary: A German consortium coordinated by the KI Bundesverband released Soofi S 30B-A3B on 2026-07-13, and it reached the Hacker News front page on 2026-07-17. It is a hybrid Mamba-2 and Transformer Mixture-of-Experts model with 31.6B total and about 3.2B active parameters per token (23 Mamba-2/MoE layers plus 6 attention layers, 128 routing experts plus 1 shared, 6 active per token). It was trained end to end on Deutsche Telekom's Industrial AI Cloud in Munich over roughly 27 trillion tokens with German up-weighted, using about 253,000 GPU-hours on Nvidia B200 hardware. The pretraining report claims the highest scores among fully open models on English and German benchmarks, ahead of OLMo 3 32B and Apertus 70B.
Why it matters: It is a European sovereign-AI open model with a released training recipe, though the weights are a gated preview checkpoint and the final license is not yet set.
Follow-up: Watch for the final license, a non-preview weight release, and independent benchmark reproduction.