A German research consortium has released Soofi S 30B-A3B, an open language model trained exclusively on Deutsche Telekom's cloud infrastructure in Munich. The model uses an efficient hybrid architecture that activates only a fraction of its 31.6 billion parameters per token, and it outperforms all fully open competitors on both German and English benchmarks.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
A German research consortium releases the fully open-source Soofi S model, which, with its hybrid architecture and German-priority training data, defeats larger competitors on multiple benchmarks. This is not only a technical breakthrough but also a key step for Europe on the path to AI sovereignty.
Soofi S adopts Nvidia's Nemotron 3 Nano hybrid architecture, combining Mamba-2 layers and standard attention layers. Of its 52 layers, only 6 require KV caches, yielding significant memory advantages: at 40K context and 32 parallel requests, Soofi S generates about 8x more tokens per GPU per second than a 14-24B parameter dense model. Similar to Alibaba's Qwen3.5 35B-A3B, its throughput remains nearly constant from 4K to 256K tokens, while traditional dense models drop sharply.
The core breakthrough of this architecture is that it demonstrates a carefully designed hybrid model can match or exceed larger models while maintaining small activated parameters. Soofi S has only 3.2B activated parameters but leads OLMo 3 32B and Apertus 70B on multiple English and German benchmarks—both of which use traditional dense architectures. However, the model scores only 56 on German competition math (Minerva MATH-DE), far behind Qwen3.5 35B-A3B (76.5) and Gemma 3 27B (65.6), possibly due to limited activated parameters restricting world knowledge storage.
Soofi S's training data comprises about 27 trillion tokens across three phases. Phase 1 uses 20 trillion general data, with 7.2% German; Phase 2 adds 6 trillion high-quality data, raising German proportion to 15.3%; Phase 3 extends context to 1M tokens. Data sources include German web text HPLT, open-source German Common Corpus, commercially licensed Genios corpus (193 million articles from 916 German publications), as well as machine-translated and synthetic German text.
This data strategy directly improves German capabilities: compared to Nvidia's Nemotron baseline, Soofi S's general German ability improves by 15.1 points, GPQA-Diamond improves by 9.6 points, and English performance remains unaffected. On the German-specific regional knowledge test INCLUDE-DE, Soofi S ties with the larger Qwen3.5 35B-A3B at 61.2 points. This shows that targeted language data allocation can effectively boost performance in a specific language without sacrificing general ability.
Soofi S is one of the first large models trained entirely on the Deutsche Telekom Industrial AI Cloud in Munich. Training took place from March to May 2026, using 512 Nvidia B200 GPUs, totaling about 253,000 GPU hours. The data center runs entirely on renewable energy, cooling water comes from the Eisbach Canal, and waste heat is used for surrounding community heating.
The project is coordinated by the German AI Association, with participants including Fraunhofer IAIS and IIS, the German Research Center for Artificial Intelligence (DFKI), Technical University of Darmstadt, and the University of Würzburg. Funding comes from the German Federal Ministry for Economic Affairs and Energy as part of the European IPCEI-CIS project. The goal is to build an open European AI model family that can run on sovereign infrastructure and be used for industrial testing. This open model avoids license restrictions of heavyweight models, providing European companies with auditable and customizable options.
Although Soofi S performs well on most benchmarks, it has clear weaknesses on certain tasks. In the RULER long-context test, when required to extract frequently appearing words from text over 32K tokens, Soofi S's hit rate plummets to about 3%, while the architecturally similar Nemotron still maintains 60-64%. The authors attribute this to a lack of synthetic data specifically for extraction tasks in the training data.
Another weakness is German competition math: Soofi S scores 56, far behind competitors. Additionally, on NaturalQuestions open factual retrieval, the model lags behind similar products. This is likely due to the model having only 3B active parameters, limiting world knowledge storage capacity. While the hybrid architecture enables efficient inference, it also brings trade-offs in knowledge capacity. Future versions may need to increase active parameters or improve training data to address these shortcomings.
This article is based on a THE DECODER report from July 13, 2026, which itself draws from the Soofi S pre-training report. All benchmark results, architecture details, and training data are from the report and were not independently verified. The source authority is moderate, but the model report itself provides verifiable data.
Soofi S demonstrates that, through hybrid architecture and language-specific data strategies, open models can challenge and even surpass much larger closed models with limited resources. It is not only a technical milestone but also an important practice for European AI sovereignty and open-source approaches. However, weaknesses in long-context retrieval and specific knowledge domains still need to be addressed in future iterations.
Primary report
Primary source