Editorial illustration for German AI Consortium's Soofi S Hits Benchmarks, Generates 8x Faster
German AI Consortium's Soofi S Hits Benchmarks,...
A German research consortium coordinated by the KI Bundesverband has released Soofi S 30B-A3B, an open-source language model trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich. The release marks one of the first large-scale training runs to run start to finish on the telecom giant's infrastructure, and the resulting model now claims the top spot among fully open systems on English and German benchmarks alike, according to its pretraining report.
What sets Soofi S apart isn't just where it was trained but how it's built. The model packs 31.6 billion parameters total, yet activates only about 3.2 billion of them for any given token, a mixture-of-experts setup borrowed from Nvidia's Nemotron 3 Nano architecture. That design choice matters for anyone running the model at scale: processing speed stays flat even as input length grows, and compute costs land closer to what you'd expect from a 3-billion-parameter model rather than a 30-billion one.
The consortium also leaned hard into German-language data during training, a bet that appears to have paid off against rivals like OLMo 3 32B and Apertus 70B. Before getting into how that architecture actually works under the hood, it's worth looking at the numbers driving the claim.
As lead author Michael Fromm writes, Soofi S positions itself between broadly multilingual European sovereignty projects like EuroLLM or Teuken, which cover many languages, and the highest-performing international open-weight models.
Why this matters
For teams paying by the GPU-hour, the throughput number here matters more than the benchmark scores. Eight times the tokens per second at 40,000 tokens of context, with flat performance as inputs grow, is the kind of claim that directly changes cost math for anyone running long-document or agentic workloads. Most 14B-24B dense models choke as context expands; if Soofi S really holds steady, that's a genuine architectural edge worth testing, not just marketing.
The 3.2-of-31.6-billion active parameter design is a concrete, checkable number, and it's worth running your own long-context benchmarks before taking Deutsche Telekom's word for it. There's also a sovereignty angle: a European consortium training entirely on domestic cloud infrastructure, with German-language performance as a stated priority, signals that the open-model race isn't only a US-China story anymore. For researchers, the interesting question is whether this hybrid-activation approach generalizes beyond German-English tasks.
For founders, the practical one is simpler: can you self-host this and actually see that 8x hold up under your own traffic patterns.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives NVIDIA Toolkit Accelerates OpenFold3 Co-Folding Workflow
From Archives Hanns Christoph Nägerl’s team finds quantum heating defies classical intuition
Further Reading
- Fraunhofer IAIS: Soofi – A Model for Industrial AI in Europe - Silicon Saxony
- Soofi announces model for industrial AI in Europe - Fraunhofer IIS
- Sovereign AI from Europe: First insights into a new model - University of Würzburg
- SOOFI: Europe's answer to ChatGPT - T-Systems
- AI for Europe – Research Field I+I - TU Darmstadt
