Soofi S: Germany Now Has an Open AI Model That Tops the Open-Source Charts
A German research consortium released Soofi S 30B, an open model trained entirely in Munich on Deutsche Telekom's cloud. It beats all fully open rivals on German and English benchmarks — and shows what 'sovereign AI' can look like in practice.
A German research consortium has released Soofi S 30B-A3B, an open language model trained entirely on Deutsche Telekom’s Industrial AI Cloud in Munich. According to its technical report, it scores higher on both German and English benchmarks than any other fully open model — including previous leaders like OLMo 3 32B from the Allen Institute and Apertus 70B from ETH Zurich.
The model has 31.6 billion parameters — roughly, the internal dials it tunes during training — but only activates about 3.2 billion of them per word it generates. This “mixture of experts” design means it runs at close to the cost of a model a tenth its size. Its hybrid architecture also keeps generation speed nearly constant even with very long inputs, where conventional models slow to a crawl. The training recipe leaned deliberately on German: up to 15.3 percent of the data mix, including a licensed corpus of 193 million newspaper articles. That focus pays off — on German benchmarks, Soofi S beats every European counterpart, sometimes by double digits. The consortium, coordinated by the German AI Association and involving Fraunhofer, DFKI, and several universities, is releasing the weights, training code, and a detailed data inventory.
What’s behind this? Europe has talked about “sovereign AI” — models built on local infrastructure, under local rules, without depending on US or Chinese providers — for years, mostly producing models that lagged far behind. Soofi S is the first German attempt that actually leads its class, at least among fully open models. Worth keeping expectations grounded, though: “fully open” is a specific league. Against open-weight heavyweights like Alibaba’s Qwen3.5 the picture is mixed, and the report is honest about weaknesses — German competition math is middling, and one long-context retrieval test drops off sharply. This is a strong 30B model, not a Fable or GPT rival.
What this means for you: If you just chat with ChatGPT or Claude, nothing changes today. But if you run models locally — or work at a company that can’t send data to US clouds — this is worth a look: a genuinely capable, genuinely open model that’s unusually strong in German, with weights free to download on Hugging Face. And for everyone in the DACH region, it’s a signal: European AI is no longer automatically the punchline.
Sources
Source: https://www.soofi.info/soofi-s/
A Turing Award Winner Bets Against Deep Learning With His New Startup Oak Lab
Richard Sutton, co-founder of modern reinforcement learning, has launched Oak Lab in Toronto. His goal: AI agents that learn continuously from experience — because he thinks today's methods are 'weak and inefficient.'
