Soofi S 30B-A3B gives Europe another serious entry in the race to build open AI systems on sovereign infrastructure. The model comes from a German research consortium coordinated by the KI Bundesverband (German AI Association), and it was trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich.
The central claim is straightforward: according to its pretraining report, Soofi S reaches the highest English and German benchmark scores among fully open models, ahead of OLMo 3 32B and Apertus 70B. The model is also built to keep generation efficient as inputs grow much longer.
A smaller active model inside a larger system
Soofi S is described as an open 30B model, but its architecture is designed so that only a fraction of the full system is active for each token. It has 31.6 billion parameters in total and activates only about 3.2 billion per generated token.
That design makes its compute profile closer to a 3B model than a conventional 30B model. The consortium uses the architecture of Nvidia's Nemotron 3 Nano without modification, combining Mamba-2 layers with standard attention layers in a hybrid mixture-of-experts system.
The long-context advantage comes from how the model handles memory. In conventional models, the KV cache used for attention grows as the context becomes longer. With many long inputs and parallel requests, that cache can become a bottleneck. Soofi S limits this pressure because only 6 of its 52 layers maintain such a cache.
The report's throughput measurements show the effect clearly. At a context length of 40,000 tokens with 32 parallel requests, Soofi S produces roughly eight times more tokens per second per GPU than dense models in the 14 to 24 billion parameter range. Its throughput stays nearly flat from 4,000 to 256,000 tokens, while conventional models slow down as context grows. Alibaba's Qwen3.5 35B-A3B, another hybrid architecture, is the only model in the measurements with similar behavior.
German data is the strategic choice
The training process covered about 27 trillion tokens across three phases. The first phase used roughly 20 trillion tokens from a broad mix of web, code, math, and domain-specific texts. A second phase used about 6 trillion tokens from higher-quality sources. A shorter third phase trained on very long documents of up to one million tokens to extend the context window.
The model's defining data decision is its emphasis on German. In the first phase, German accounts for 7.2 percent of the training mix. In the second phase, that rises to 15.3 percent. By comparison, Nvidia's Nemotron reference recipe assigns only about 5 percent to all non-English languages combined.
The German sources include web text from HPLT, the openly licensed German Commons corpus, German portions of FinePDFs and FineWiki, and the commercially licensed Genios corpus. Genios contributes 193 million newspaper articles from 916 German publications. Machine-translated and synthetically generated German texts are also part of the mix.
Where Soofi S leads and where it falls short
In evaluations against 16 other open models, Soofi S leads all fully open models on aggregate German and English scores, according to the report. That comparison includes OLMo 3 32B from the Allen Institute for AI and Apertus 70B from ETH Zurich and EPFL.
The model also ranks strongly against European sovereign baselines. On all German benchmarks in the suite, it comes out ahead, sometimes by double-digit margins.
Its coding results are also notable among open-source peers:
- 73.8 percent on HumanEval
- 70.2 on MBPP
- 84.2 on the German MBPP variant
On INCLUDE-DE, a benchmark for Germany-specific regional knowledge, Soofi S ties for first place at 61.2 points with the larger Qwen3.5 35B-A3B. Compared with the Nemotron baseline, its German data recipe improves language proficiency by 15.1 points and GPQA-Diamond by 9.6 points, while the report says English performance is not sacrificed.
The model is not strongest everywhere. On German competition math, Soofi S scores 56 points on Minerva MATH-DE, behind Qwen3.5 35B-A3B at 76.5 and Gemma 3 27B at 65.6. It also trails on open factual retrieval in NaturalQuestions, which the source links to the limits of having only 3 billion active parameters and therefore less room to store world knowledge than a dense 27B model.
The RULER long-context test exposes another specific weakness. When asked to extract frequently occurring words from a long text, Soofi S's hit rate falls to around 3 percent beyond 32,000 tokens of context. The comparable Nemotron model reaches 60 to 64 percent on that task. The authors attribute the gap to long-context training data that includes many long documents but lacks synthetic data designed for extraction tasks. On the remaining twelve RULER tasks, the two models perform about the same.
Sovereign infrastructure and open release details
The training run took place between March and May on up to 512 Nvidia B200 GPUs at Deutsche Telekom's Industrial AI Cloud in Munich. The total was about 253,000 GPU-hours. According to the report, the facility runs entirely on renewable energy, uses water from the Eisbach canal for cooling, and feeds waste heat into the surrounding Tucherpark neighborhood.
The consortium includes German research institutions and companies. Participants named in the source include the Fraunhofer Institutes IAIS and IIS, the German Research Center for Artificial Intelligence (DFKI), TU Darmstadt, the University of Würzburg, the L3S Research Center, the Berlin University of Applied Sciences, Ellamind, and Merantix Momentum. The project is funded by the German Federal Ministry for Economic Affairs and Energy as part of the European IPCEI-CIS program.
The release is meant to be open in a documented way. The researchers are releasing model weights with selected intermediate checkpoints, the complete training and evaluation code, and a detailed data inventory with raw token counts, epoch numbers, and effective contributions per source. They also document sources that were reviewed but excluded.
According to the team, Soofi S meets the Open Source AI Definition 1.0 from the Open Source Initiative. A stricter European open-data proposal would not be met because 1.3 percent of the training mix comes from commercially licensed Genios data. The report says about 99 percent of the training mix can be independently reconstructed, while the exact license for the model's release has not been finalized yet.
The project positions Soofi S between broad multilingual European sovereignty efforts such as EuroLLM or Teuken and the strongest international open-weight models. The consortium is now looking for industry partners for a next phase focused on applications involving technical documents, code generation, and agent-based systems.