A consortium of German research institutions and AI companies has released Soofi S 30B-A3B, an open language model that the project says achieves the highest scores on both English and German benchmarks among fully open models. Coordinated by the KI Bundesverband, Germany's national AI association, the release is one of the first large language models trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich and is positioned as a sovereign European alternative to the dominant United States and Chinese open-weight releases. For developers tracking the latest AI developments, the launch is notable less for raw scale than for the unusual efficiency and language focus baked into the architecture.
A hybrid design with only 3.2 billion active parameters
Soofi S is a mixture-of-experts (MoE) model with 31.6 billion parameters in total, but it activates only about 3.2 billion per generated token. That puts its compute cost closer to a 3-billion-parameter model than a conventional 30-billion-parameter dense model, a design choice aimed at keeping inference cheap enough to run on European infrastructure without reliance on overseas cloud providers.
The architecture is borrowed from Nvidia's Nemotron 3 Nano, combining Mamba-2 layers with standard attention layers in a hybrid layout. According to the project's pretraining report, only 6 of the model's 52 layers maintain a key-value (KV) cache — the memory structure that stores previous tokens for attention computation. In conventional transformers, that cache grows linearly with context length and becomes a major bottleneck during long-context inference.
The practical payoff shows up in throughput. 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 generation speed for conventional models drops sharply as context grows, Soofi S stays nearly flat from 4,000 tokens all the way to 256,000 tokens. The only comparable model in the consortium's measurements is Alibaba's Qwen3.5 35B-A3B, which also uses a hybrid architecture.
Trained for German, without sacrificing English
The deliberate focus on German is central to the project. In the first training phase, German makes up 7.2 percent of the data mix; in the second phase, that share rises to 15.3 percent. By comparison, in Nvidia's Nemotron reference recipe all non-English languages combined account for only around 5 percent of the training mix.
Data sources include German web text from the HPLT corpus, the openly licensed German Commons corpus, German portions of FinePDFs and FineWiki, and the commercially licensed Genios archive of 193 million newspaper articles drawn from 916 German publications. Machine-translated and synthetically generated German texts round out the mix.
The model processed roughly 27 trillion tokens across three training phases: about 20 trillion tokens of broad web, code, math, and domain-specific text in the first phase; around 6 trillion higher-quality tokens in the second; and a shorter third phase extending the context window using documents up to one million tokens long.
Benchmark results: ahead on German and English, weaker on math
In evaluations against 16 other open models, Soofi S leads all fully open models on aggregate scores for both German and English, according to the consortium. That includes OLMo 3 32B from the Allen Institute for AI and Apertus 70B from ETH Zurich and EPFL. Against every European sovereign baseline, the model comes out ahead on all German benchmarks in the suite, sometimes by double-digit margins.
On code tasks, Soofi S scores 73.8 percent on HumanEval, 70.2 on MBPP, and 84.2 on the German MBPP variant — the best results among open-source peers. On INCLUDE-DE, a test 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, the German-weighted data recipe improves language proficiency by 15.1 points and the GPQA-Diamond science benchmark by 9.6 points, without hurting English performance.
There are clear weaknesses. On German competition mathematics (Minerva MATH-DE), Soofi S scores 56 points, well 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 team attributes to having only about 3 billion active parameters and therefore less stored world knowledge than a dense 27B model. The RULER long-context test surfaces another gap: when the model has to extract frequently occurring words from a long text, its hit rate drops to around 3 percent beyond 32,000 tokens, while the comparable Nemotron model still manages 60 to 64 percent.
Trained on 512 Nvidia B200 GPUs in Munich
The training run took place between March and May 2026 on up to 512 Nvidia B200 GPUs at Deutsche Telekom's Industrial AI Cloud in Munich, totaling about 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, according to the report. Soofi S was one of the first major training runs carried out on this infrastructure.
The consortium behind the model is funded by the German Federal Ministry for Economic Affairs and Energy as part of the European IPCEI-CIS program. Participants 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, and AI companies Ellamind and Merantix Momentum.
A debate over 'overtraining'
Soon after launch, critics argued that Soofi S was heavily overtrained by the standards of the classic Chinchilla scaling laws published by Google DeepMind in 2022, which suggested a rough sweet spot of 20 tokens per parameter. With 27 trillion tokens and roughly 30 billion parameters, Soofi S lands at several hundred tokens per parameter; counting only the 3.2 billion active parameters pushes the ratio into the thousands.
Michael Fromm, part of the project's technical leadership, pushed back on that criticism, arguing that those rules do not carry over to MoE architectures. "There's new research showing that the old scaling laws from dense models no longer apply to MoE architectures," Fromm said, noting that individual experts benefit from seeing the same documents repeatedly in a large, high-quality dataset. As a point of comparison, he pointed to Nvidia, which has trained its own models on up to 25 trillion tokens.
What gets released, and what does not
The researchers are releasing model weights along with selected intermediate checkpoints, the complete training and evaluation code, and a detailed data inventory listing raw token counts, epoch numbers, and effective contributions per source. According to the team, this means Soofi S meets the Open Source AI Definition 1.0 from the Open Source Initiative.
A stricter proposal for a European open-data definition, which would require every single training token to be freely distributable, is not met because 1.3 percent of the mix comes from the commercially licensed Genios corpus. The report says about 99 percent of the training data can still be independently reconstructed. The exact license for the model's release had not been finalized at the time of publication.
The consortium is now looking for industry partners for the next phase to test Soofi S in applications involving technical documents, code generation, and agent-based systems. For more on open-weight model releases, sovereign AI infrastructure, and the European AI ecosystem, keep up with the AI industry coverage published here.
Stay ahead on open-source AI
Open-weight releases are arriving almost weekly from labs in the United States, China, and now Europe, and the gap between the largest proprietary models and the best open alternatives keeps narrowing. Follow the breaking AI news and benchmark analysis here for the full picture.
Read more AI model coverage →
