What happens when a consortium of German research institutions decides to build Europe's answer to open-source AI, trains it entirely on a sovereign cloud running on renewable energy, and publishes every scrap of training data documentation alongside the weights? You get Soofi S, a 31.6 billion parameter Mixture-of-Experts model that activates only 3.2 billion parameters per token, was trained on 27 trillion tokens across three phases, and tops every fully open model in both English and German benchmarks. The model, released July 13, 2026, by a consortium coordinated by the KI Bundesverband (German AI Association) and funded by the German Federal Ministry for Economic Affairs, is the strongest signal yet that European AI sovereignty is moving from policy papers to production.
Soofi S is not just another open model release. It represents a deliberate architectural bet, a geopolitical statement, and a transparency commitment that most model releases still avoid. For founders evaluating which open models to build on, this release introduces a new tier of European option: one that is GDPR-compatible by design, runs on known infrastructure, and publishes the full data inventory including what was excluded and why.
A Hybrid Architecture Built for Long Contexts
Soofi S follows the architecture of Nvidia's Nemotron 3 Nano, combining Mamba-2 state-space layers with standard attention layers. The model has 52 layers total, but only six maintain a KV cache for attention computation. The rest use Mamba-2's linear-time state-space mechanism, which eliminates the quadratic memory scaling that plagues dense transformers at long context lengths. The practical effect is dramatic: at a context length of 40,000 tokens with 32 parallel requests, Soofi S generates roughly eight times more tokens per second per GPU than dense models in the 14 to 24 billion parameter range. While conventional models show throughput dropping sharply as context grows, Soofi S stays nearly flat from 4,000 to 256,000 tokens. The only other model that shows this behavior in published benchmarks is Alibaba's Qwen3.5 35B-A3B, which also uses a hybrid architecture. For solo founders running inference on constrained hardware, this matters. A model that maintains throughput at 256K context on consumer GPUs is a model that can handle full codebase analysis, long document processing, and multi-turn agentic workflows without the costly context window reshuffling that plagues dense models at scale.
Benchmark Performance That Punches Above Its Weight Class
Soofi S achieves the highest aggregate scores among fully open models across both English and German benchmark suites. On coding benchmarks, it scores 73.8 percent on HumanEval, 70.2 on MBPP, and 84.2 on the German-language variant of MBPP. On INCLUDE-DE, a benchmark for Germany-specific regional knowledge, it ties for first place at 61.2 points with the larger Qwen3.5 35B-A3B. Against European sovereign baselines specifically, Soofi S leads every German-language benchmark, sometimes by double-digit margins. These results put it ahead of OLMo 3 32B from the Allen Institute for AI and Apertus 70B from ETH Zurich. Being a 3.2 billion active parameter model beating 32B and 70B dense models is noteworthy and validates the efficiency-first approach that MoE advocates have been pushing for the past two years. There are real weaknesses. On German competition math (Minerva MATH-DE), Soofi S scores 56 points versus Qwen3.5's 76.5 and Gemma 3's 65.6. On open factual retrieval (NaturalQuestions), its smaller active parameter count limits world knowledge storage. And on the RULER long-context test, when asked to extract frequently occurring words from texts beyond 32,000 tokens, its hit rate drops to roughly 3 percent versus the Nemotron baseline's 60 to 64 percent. The authors attribute this to a lack of synthetic extraction training data, which is fixable in future iterations.
Sovereign Infrastructure and the Overtraining Debate
The training ran on up to 512 Nvidia B200 GPUs at Deutsche Telekom's Industrial AI Cloud in Munich, totaling approximately 253,000 GPU-hours. The facility runs entirely on renewable energy, is cooled with water from the Eisbach canal, and feeds waste heat into the surrounding Tucherpark neighborhood. That is a level of infrastructure transparency that most hyperscaler-based training runs do not provide. The consortium published model weights, selected intermediate checkpoints, complete training and evaluation code, and a detailed data inventory listing raw token counts, epoch numbers, and effective contributions per source. Sources reviewed but excluded are also documented. This means Soofi S meets the Open Source AI Definition 1.0 from the Open Source Initiative. The consortium acknowledges that it does not meet stricter open-data proposals that would require every training token to be freely distributable, because 1.3 percent of the data comes from the commercially licensed Genios corpus containing 193 million newspaper articles from 916 German publications. The model also faces an overtraining controversy. Critics argue that with roughly 27 trillion tokens across 30 billion parameters, Soofi S far exceeds the classic Chinchilla scaling law's recommended ratio of approximately 20 tokens per parameter. Michael Fromm, part of the project's technical leadership, pushes back: the old scaling laws from dense models simply do not carry over to MoE architectures, where individual experts benefit from seeing the same documents multiple times. He points to Nvidia training its own models on up to 25 trillion tokens as precedent.
What This Means for Builders and European AI Independence
For founders building AI products, Soofi S offers three immediate practical advantages. First, a GDPR-compatible open model trained on European infrastructure eliminates the legal uncertainty that comes with deploying US or Chinese models in regulated European enterprise environments. Second, the hybrid architecture delivers production-grade throughput at long context lengths on modest hardware, which directly reduces inference costs for document-heavy workflows. Third, the transparency commitment means you can audit exactly what went into the model, which matters for compliance-sensitive applications in healthcare, legal, and finance. The broader implication is that European AI sovereignty is no longer theoretical. The German government backed this through the IPCEI-CIS program, Deutsche Telekom provided the compute, and a consortium of top research institutions executed. The playbook is replicable: government funding plus sovereign cloud infrastructure plus academic research talent equals a competitive open model. For founders operating in Europe, the arrival of Soofi S means you now have a credible open-source alternative that can be deployed without triggering GDPR cross-border transfer concerns, while matching or beating US and Chinese open models on the benchmarks that matter most for European use cases.