German AI consortium releases open language model Soofi S with focus on German and English

A German research consortium released Soofi S 30B-A3B on 13 July 2026, an open language model that scores higher than all fully open alternatives on both German-language and English-language benchmarks. The model has 31.6 billion parameters and was trained from November 2025 to May 2026 on Deutsche Telekom's cloud infrastructure in Munich.

The project is coordinated by KI Bundesverband, the German industry association for artificial intelligence, and is funded by the German Federal Ministry for Economic Affairs and Energy as part of the European IPCEI-CIS programme. The ministry's contribution amounts to approximately twenty million euros. The consortium includes, among others, the Fraunhofer Institutes IAIS and IIS, the German Research Center for Artificial Intelligence (DFKI), TU Darmstadt, the University of Würzburg, and AI companies Ellamind and Merantix Momentum.

The actual training run ran from 24 March to 13 May 2026. On 9 July the accompanying paper appeared on arXiv; four days later the model weights were made available, initially in a closed beta phase for industrial partners.

Architecture combines Mamba and Transformer

Soofi S uses a hybrid Mixture-of-Experts architecture that combines Mamba layers with traditional Transformer layers. Of the 31.6 billion parameters, only approximately 3.2 billion are activated per token, which limits the computational load per inference step.

That design offers measurable advantages with long inputs. At a context length of 40,000 tokens, the model achieves roughly eight times the generation throughput of comparable dense models. The supported context window extends to one million tokens, enabling deployment in applications involving large documents or lengthy conversation histories.

The model was trained on approximately 27 trillion tokens, with a deliberate overrepresentation of German-language material. Technical project lead Dr. Nicolas Flores-Herr is affiliated with Fraunhofer IAIS; the broader research team also includes researchers from DFKI, the University of Würzburg, and the L3S Research Center in Hannover.

Training infrastructure from Deutsche Telekom and NVIDIA

Training took place entirely on Deutsche Telekom's Industrial AI Cloud in Munich. That facility has more than ten thousand GPUs and was built through a one-billion-euro partnership between Deutsche Telekom and NVIDIA. Leibniz University Hannover awarded the associated infrastructure contract to Deutsche Telekom.

The choice of this infrastructure was deliberate: the consortium wanted to train the model outside American cloud services, as part of broader European efforts around digital sovereignty. Jörg Bienert, managing director of both the Center for Sovereign AI and KI Bundesverband, coordinates the overarching project.

Benchmark position and availability

According to the initial results presented on 17 June 2026, Soofi S leads all fully open models on aggregated benchmark scores for both German and English. The arXiv paper contains the detailed comparisons; the exact scores per benchmark can be found there.

The model is currently in a closed beta phase. The model repositories are for now only accessible to industrial partners testing the model in scenarios involving technical documentation, code generation, and agent-based systems. Once the beta phase concludes, the weights will be made freely available. The consortium is actively seeking additional industrial partners for this testing phase.

Significance for the European AI ecosystem

Soofi S is an example of how European public-private consortia are attempting to close the gap with American and Chinese frontrunners in the area of open, large-scale language models. The combination of government funding via IPCEI-CIS, academic capacity from multiple universities and institutes, and private cloud infrastructure is an approach comparable to initiatives such as the French Lucie model and the pan-European OpenGPT-X project.

For Dutch and European founders and investors, the release is relevant as a reference point: a model trained entirely in Europe that competes on benchmarks with globally available alternatives. The choice of a gated beta phase with industrial partners also offers insight into how these types of consortia seek to accelerate commercial applications without immediately relinquishing the model's openness. Whether that approach yields broad adoption will depend on the experiences industrial partners gain in the months ahead.