For years, the AI narrative has been a two-horse race. The US builds frontier models. China scales them fast. Europe watches from the sidelines, worrying about sovereignty but doing little about it. That changed on July 10. A German research consortium released Soofi S 30B-A3B, the first large-scale language model trained entirely on European infrastructure — Deutsche Telekom's Industrial AI Cloud in Munich.
It is not the smartest model in the world, and it is not meant to be. It is a different kind of statement.
Soofi S is a Mixture-of-Experts model with 31.6 billion total parameters, but it activates only 3.2 billion per token — putting its inference cost closer to a 3B model than a 30B one. The hybrid Mamba-Transformer architecture keeps throughput flat from 4,000 to 256,000 tokens of context. At 40,000 tokens with 32 parallel requests, it generates roughly eight times more tokens per second per GPU than dense models in the 14-24B range.
This is not a general-purpose chatbot. It is a foundation model for industrial AI — document analysis, code generation, agentic workflows, and German-language applications. It is designed to run on sovereign infrastructure, not just be trained on it.
The German focus is deliberate. In the first training phase, German made up 7.2 percent of the data mix; in the second phase, that rose to 15.3 percent. For comparison, Nvidia's Nemotron reference recipe allocates only about 5 percent to all non-English languages combined. The model now leads all fully open models on both English and German benchmarks, surpassing OLMo 3 32B and Apertus 70B. On code generation, it scored 73.8 percent on HumanEval and 70.2 on MBPP — the best among open-source peers.
The project is transparent by design. The consortium is releasing weights, intermediate checkpoints, training code, and detailed data accounting. This is not a "weight-only" release. It is a fully auditable model — a requirement for European enterprises that cannot risk vendor lock-in or opaque training practices.
But there are limits. On German competition math, Soofi S scored 56 points on Minerva MATH-DE, well behind Qwen3.5 35B-A3B (76.5) and Gemma 3 27B (65.6). On the RULER common-word extraction task beyond 32,000 tokens, its hit rate dropped to around 3 percent. The model outperforms its European peers, but it is not chasing the global frontier. It is chasing a different prize: independence.
Europe does not need to build the smartest AI. It needs to build AI it controls. Soofi S is the first serious proof that it can. Europe has spent years talking about digital sovereignty — about reducing dependence on American and Chinese AI. Soofi S is the first time a European model has been trained end-to-end on European infrastructure, with European data, for European use cases. It will be tested in industrial applications involving technical documents, code generation, and agent-based systems — not as a trophy, but as a tool.
P.S. If you are an enterprise AI buyer in Europe, you now have a local option that meets GDPR requirements, runs on European hardware, and is fully auditable. It may not beat GPT-5.6 on every benchmark, but it runs on your terms — and sometimes that matters more than winning the leaderboard.
