In brief: Soofi S is an open German language model with efficient mixture-of-experts architecture that outperforms European and U.S. open-source competitors in benchmarks and was trained on Deutsche Telekom infrastructure.
The Soofi Research Consortium has released the open-source model Soofi S 30B-A3B, a hybrid Mamba-Transformer model for German and English that outperforms European and U.S. open-source competitors in benchmarks. The model is based on a mixture-of-experts architecture and was trained entirely on Deutsche Telekom infrastructure.
Soofi S 30B-A3B is a 30-billion-parameter model that activates only 3 billion parameters per token. This mixture-of-experts architecture significantly reduces computational requirements and keeps memory usage nearly constant even with longer text sequences. The model is therefore particularly well-suited for high-frequency multi-user scenarios and processing extensive contexts.
In technical benchmarks for English and German, Soofi S outperforms models such as Alia 40B, EuroLLM 22B, Apertus 70B, and Olmo 3 32B. In aggregated tests, it achieves the performance of closed models with 14 to 27 billion parameters at significantly lower active parameter costs. Michael Fromm, lead of pretraining data at Soofi, summarizes the approach: “Our advantage is performance + throughput + complete openness.” The model performs more weakly on knowledge tests (MMLU-Pro) and advanced scientific reasoning (GPQA-D-DE), but remains ahead of other open-source competitors in these areas.
The model is not available as a publicly accessible chatbot. Companies and organizations must submit an application with a specific use case. Politicians such as Karl Lauterbach (Committee on Research, Technology and Space) do not view Soofi S as competition to leading models like Claude or OpenAI systems, but see practical value for industrial and administrative applications.
Source: www.it-daily.net · Published 17 July 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.7.3.