Soofi S: Germany's first serious AI model – The safe AI solution for SMEs?

Germany's new AI model Soofi S: A real breakthrough or just "good for Europe"?

Soofi S review: How does the new German language model fare against the global AI elite?

For a long time, the race for technological supremacy in the field of artificial intelligence seemed to have been decided – fought exclusively between US tech giants and state-subsidized Chinese initiatives. Europe risked being relegated to the role of mere consumer and regulator. But now, Germany's AI sector is making a triumphant return to the international stage: The public-private consortium behind the SOOFI project is presenting "Soofi S 30B-A3B," a language model that is among the world's leading fully open systems.

Trained on local infrastructure in Munich and designed with a radical focus on absolute data transparency and GDPR compliance, it aims to offer a sovereign alternative, especially for small and medium-sized enterprises (SMEs) and highly regulated industries. But does the model stand up to harsh reality? A closer look at the benchmark results, the innovative hybrid architecture, and the sobering market reality reveals that Soofi S is a remarkable milestone and proof that Europe can build competitive AI – but it is far from the end of a long, arduous road to true digital independence. A comprehensive analysis.

Between benchmark fame and frontier reality – why “good for Europe” is not a sufficient answer

The German AI consortium has released Soofi S 30B-A3B, a language model that leads the world among fully open models – yet still lags behind the Chinese Qwen3.5. This simultaneous occurrence of genuine progress and sobering relativization is key to understanding what is currently happening in the German AI landscape.

What makes Soofi S technically special

The model bears the official designation 30B-A3B, which precisely describes its architecture: 31.6 billion parameters in total, but only around 3.2 billion of them are active per processed token. This discrepancy is not a flaw, but rather the core of an intelligent architectural principle. Soofi S relies on a hybrid Mixture of Experts structure that combines Mamba 2 layers with classic Transformer Attention layers – a concept that the consortium directly adopted from Nvidia's Nemotron 3 Nano and further developed.

The advantages of this architecture only become apparent under real-world conditions. While dense models require increasingly more computing power with growing context length, resulting in a significant drop in throughput, Soofi S remains almost constantly efficient. With a context length of 40,000 tokens and 32 concurrent requests, it generates roughly eight times more tokens per second per GPU than comparable dense models with between 14 and 24 billion parameters. Only 6 of the 52 layers maintain a kv cache, which keeps memory pressure low even with very long documents. The context window extends up to one million tokens – a size that makes applications with massive document volumes or lengthy conversation histories practically feasible.

The actual computing effort of the training, which ran between March 24 and May 13, 2026, on up to 512 NVIDIA B200 cards in Deutsche Telekom's Industrial AI Cloud in Munich, totaled 253,000 GPU hours. According to the project report, the facility uses entirely renewable electricity, is cooled with water from the Eisbach stream, and feeds the waste heat back into the Tucherpark industrial park – a detail that, in an industry with exorbitant energy demands, is more than just eco-marketing.

How training re-evaluates the German language

The training corpus comprises approximately 27 trillion tokens – a dataset that truly rivals Frontier's offerings and explains the significant qualitative leap compared to previous European attempts. Anyone wanting to understand why predecessors like Apertus, EuroLLM, Teuken, and Salamandra lagged so far behind international standards in benchmark comparisons will find the clearest answer here: they simply trained with too little data. Scalability and data volume are not optional luxuries in language model development, but rather crucial prerequisites for performance.

Within this corpus, the consortium deliberately overemphasized the German language. In the first training phase, German accounts for 7.2 percent of the total training mix, and in the second phase, this share increases to 15.3 percent. By comparison, in Nvidia's Nemotron recipe, all non-English languages ​​combined account for approximately 5 percent. This deliberate bias explains why the model performs so well on German benchmarks.

The data sources are unusually transparently documented. In addition to HPLT web texts and the German Commons corpus, a commercially licensed Genios database containing 193 million newspaper articles from 916 German publications was incorporated into the training. According to the consortium, around 99 percent of the entire training mix is ​​traceable and publicly accessible – which represents a paradigm shift in an industry where even large US companies treat training data as trade secrets. This includes selected intermediate states of the model, hyperparameters, complete training code, and evaluation code.

Where Soofi S stands in the benchmark field

A sober assessment requires reconciling two truths. On the one hand, according to the consortium report, Soofi S leads all fully open models in an aggregated German benchmark score with 79.1 points – ahead of Olmo 3 32B from the Allen Institute and Apertus 70B from Switzerland. In English-language benchmarks, the model is also the strongest among the fully open alternatives. For coding tasks, it achieves 73.8 percent on HumanEval and 70.2 percent on MBPP.

On the other hand, this leading field is a subcategory, not a global ranking. Qwen3.5 35B-A3B, Alibaba's Chinese model, achieves 76.5 points in German-language competitive mathematics, while Soofi S scores 56 points. This is not a marginal deficit, but a substantial gap precisely where abstract reasoning is required. Soofi S also falls behind in international comparisons with models like Qwen3.6 27B or GLM 5.2, and these competitors are rightly considered benchmarks in the professional community.

The benchmarks themselves are also subject to critical scrutiny. Jenia Jitsev from the LAION consortium described the consortium's self-defined capability index metric as overstated. And a data mining professor raised the crucial question of whether the presented figures were independently evaluated or whether they were simply self-reported data that had not been independently reproduced. This methodological skepticism is justified and cannot be dismissed: Benchmark results only gain credibility through independent reproduction, not through self-reporting.

The consortium and the infrastructure behind it

Soofi is not a private startup project, but a public-private consortium project that Germany has embedded within a European framework. It is coordinated by the German AI Association, the German industry association for artificial intelligence. The federal government has provided approximately €20 million in funding through the Federal Ministry for Economic Affairs and Climate Action, within the European IPCEI-CIS framework. The acronym SOOFI stands for "Sovereign Open Source Foundation Models for European Intelligence"—the name itself is programmatic.

On the research side, the consortium boasts remarkable institutional depth: Fraunhofer IAIS and Fraunhofer IIS, the German Research Center for Artificial Intelligence (DFKI), TU Darmstadt, the University of Würzburg, Leibniz University Hannover, and the L3S Research Center contribute the academic expertise. The AI ​​companies Ellamind and Merantix Momentum are participating from the industry. Dr. Nicolas Flores-Herr from Fraunhofer IAIS is responsible for the technical project management.

The underlying infrastructure is the result of a billion-euro partnership between Deutsche Telekom and NVIDIA: The Industrial AI Cloud in Munich operates over ten thousand GPUs, including, from March 2026, a network of approximately 130 NVIDIA DGX B200 systems with a total of over 1,000 GPUs, which will be used exclusively for European language modeling projects. The contract for this infrastructure was awarded to Telekom via Leibniz University Hannover – a process deliberately located in Germany with a clear rationale: no training on American cloud infrastructure.

What true openness means – and why it matters

The term "open source" has become overused and often misleading in the AI ​​industry. Many models are marketed as "open" even though only the finished weights are available for download – without training data, without code, and without insight into the data composition. This form of openness is sufficient for everyday business use, but it doesn't create genuine control or allow for independent verification.

Soofi S goes further structurally. The publication includes model weights, selected training checkpoints, the complete training code, all evaluation scripts, and a complete breakdown of the training data sources with precise mixing statistics. Where source data is under permissive licenses, the construction artifacts are also released; commercially licensed sources are documented with aggregate statistics. These are the prerequisites that regulated industries need for auditability and that the EU AI Act will require in the future anyway.

For sectors like financial services, medical technology, or public administration, this traceability is not merely an aesthetic advantage, but a legal requirement. A bank or insurer using an AI model in an auditable process must be able to document what data has been fed into the model and who retains technical control over it. US-based Frontier models cannot structurally answer this question—not because they are unwilling to, but because the training data is considered a core trade secret.

This strength is limited by one unresolved issue: the final commercial license is still pending at the time of release. Anyone planning a production deployment today must wait for this matter to be resolved. This is a real obstacle for early adopters and should be omitted from any honest assessment.

The argument of digital sovereignty

The question of whether "sovereign AI" is more than just a buzzword can be answered concretely for the first time with Soofi S – at least partially. Training on German infrastructure, outside of American clouds, is not merely symbolic: it prevents NVIDIA or hyperscaler terms and conditions from being applied to training data and avoids the extraterritorial reach of the US Cloud Act, which, in principle, grants US authorities access to data processed on US infrastructure, regardless of the server location.

For many companies based in Germany, this control is a real and business-relevant issue. Those operating a language model containing internal design plans, confidential customer data, or medical information face a fundamental trust problem with US services – not out of paranoia, but due to risks that are not fully clarified legally. A model that runs entirely on German servers, has fully documented training data, and is permissively licensed structurally eliminates this legal gray area.

The KPMG study on the AI ​​Geopolitics Index 2026 confirms the structural framework: Europe achieves only 48.8 points in the Strategic AI Capability Index, compared to 75.2 for the USA. The DACH region, with 54 points, is slightly below Western Europe and struggles with fragmented capital markets, high energy prices, and limited computing capacity for growth companies. In this context, Soofi S is not a breakthrough in itself, but it is a concrete counterweight to the otherwise complete technological dependence on non-European providers.

Our EU and German expertise in business development, sales and marketing - Image: Xpert.Digital

Industry focus areas: B2B, digitalization (from AI to XR), mechanical engineering, logistics, renewable energies and industry

More information here:

• Expert Business Hub

A thematic hub offering insights and expertise:

• Knowledge platform covering global and regional economies, innovation and industry-specific trends
• A collection of analyses, insights, and background information from our key areas of focus
• A place for expertise and information on current developments in business and technology
• A hub for companies seeking information on markets, digitalization, and industry innovations

Where the model finds its place – and where it doesn't

The debate surrounding Soofi S risks conflating two fundamentally different questions: Is it a frontier model that competes with GPT-5 or Gemini 2.5? And is it a useful, practically applicable tool for specific use cases? The first question can clearly be answered with a no. The second is more complex.

For complex reasoning tasks, large-scale software development, in-depth scientific analysis, or large-scale creative projects, Soofi S falls short of the major proprietary models. Those seeking the best available AI assistant for demanding generative tasks will currently be better served by Qwen3.5, Claude, or GPT-5. This finding is neither surprising nor a disgrace—it is the logical consequence of the resource disparity between a 20-million-euro consortium research project and the multi-billion-dollar US and Chinese AI labs.

The picture is quite different where the model is actually intended for use: in industrial processes, in German public administration, on edge hardware in production environments, or on company servers with GDPR requirements. Soofi S was explicitly designed for precisely this area of ​​application. Real-time machine monitoring, quality control, operator assistance on the production line, compliance pre-checks, ticket triage, local fault diagnosis on CNC machines, predictive maintenance alerts – these are tasks where a model with 3.2 billion active parameters and constant memory requirements over long contexts offers structural advantages. For these scenarios, latency is more important than eloquence, and throughput is more important than literary richness.

The mixture-of-experts architecture with consistently low KV cache requirements is optimized for these scenarios. With 40,000 tokens of context and 32 parallel queries, Soofi S outperforms dense models by a factor of eight in throughput. This is not an abstract academic benchmark, but a key performance indicator that determines the cost-effectiveness of a local, on-premises deployment.

The middle class as the actual target group

In the consortium's press release, Soofi S is explicitly described as a model for SMEs – and this positioning is more consistent than it initially appears. Small and medium-sized enterprises (SMEs) in Germany face a specific set of challenges: They typically lack dedicated machine learning teams capable of fine-tuning proprietary frontier models. They often process sensitive customer data or trade secrets, for which cloud-based US models are problematic due to compliance concerns. And they seek solutions that are locally operable, documentable, and manageable during operation.

For this profile, a permissively licensed, fully transparent, medium-sized model with strong German language skills is indeed more attractive than a higher-performing model whose training data, weights, and licensing structure remain opaque. Bitkom figures support this assessment: Two-thirds of Germans express a desire to use AI from Germany – this is not a technical preference, but rather a preference for data privacy and trust, which is reflected in procurement processes and customer requirements.

At the same time, medium-sized businesses are not a homogeneous category. An automotive supplier with global supply chains, English-language communication, and complex design tasks faces different requirements than a regional administrative authority or a law firm with confidential correspondence. The first group will not find a complete solution in Soofi S. The second group, however, might discover in it a valuable core component of a sovereign AI stack.

What the model reveals about Germany as an AI location

The Expert Commission on Research and Innovation (EFI) painted a sobering picture in its 2026 annual report: strong basic research, but hardly any proprietary models, insufficient computing capacity, and a GDPR that hampers European developers while US models operate unhindered in the EU market. Soofi S is a direct response to precisely this diagnosis – and simultaneously its best proof that change is possible.

The PwC AI Fitness Index 2026 ranking attests to Germany's strength in governance and data, but this strength doesn't translate into business impact. This is precisely the core problem: Germany excels at regulation and documentation, but struggles with scaling and commercialization. Soofi S replicates this pattern: full transparency, a clear compliance architecture, academic depth – but no marketable product that will be running in a mid-sized company's production line tomorrow. At the time of publication, the model is still in closed beta, accessible only to select industry partners.

The acquisition of Aleph Alpha by Cohere in April 2026 is revealing in this context. It demonstrates an alternative approach: instead of building their own top-tier platform, some providers rely on sovereign operation and compliance layers built on top of foreign models. This approach is more realistic for many mid-sized companies than waiting for a consortium model. However, it doesn't completely solve the sovereignty problem – it merely shifts it to the operator level.

What's missing between research project and market product

One of the most productive misunderstandings surrounding Soofi S is the confusion between research success and market success. The consortium of Fraunhofer, DFKI, universities, and startups has indeed achieved something no one in Europe has managed before: training a language model at the frontier data level with complete transparency and a European infrastructure. The fact that this required a consortium of research institutions rather than profit-driven private companies is not a sign of strength, but rather an indication of a structural weakness in the European AI ecosystem.

Market readiness is not a given. A model needs functioning licenses, production stability, deployment tools, support structures, fine-tuning pipelines, and integrable APIs before it can truly be used in an enterprise. The final license is still pending at the time of publication. The model is in closed beta with industry partners who are testing it for technical documentation, code generation, and agent-based systems. This is the right step, but it underscores how far there is still to go from an impressive research result to a production-ready enterprise tool.

In addition, there is the licensing issue for the training model itself. A comment from the expert community points to the different variants within the model family – Isar and Rhine – and warns against starting to use it before the commercial licensing issue is definitively resolved. This caution is justified, because a model that is integrated into critical business processes and later proves to be non-commercially usable will generate considerable technical and legal costs for reversing the process.

The real benchmark: scalability and ecosystem

What ultimately becomes of Soofi S depends less on the quality of the current model than on the consortium's and the German AI landscape's ability to build upon it. The project has explicitly announced a family of models, not just a single one. The initial goal of 100 billion parameters was communicated in December 2025 – Soofi S, with its 30 billion, is the first building block.

If this initial building block evolves into a complete model family that is regularly updated, scales with Telekom's computing infrastructure, and attracts a genuine industrial ecosystem of fine-tuning providers, integrators, and application manufacturers, then that will be a true breakthrough. If it remains a proof of concept—an academic success without commercial success—then Soofi S will join a long list of European projects that began with great fanfare and fizzled out in operation.

The decisive indicators for future developments are therefore not today's benchmarks, but rather the speed of licensing, the breadth of beta partners and their public feedback, whether a follow-up project for the larger model is already funded, and finally, whether private companies with a profit motive participate in further development or whether the model remains permanently dependent on public funding. AI sovereignty is not achieved through labels, but through performance, scalability, and a market that allows and rewards innovation.

European context and geopolitical dimension

Soofi S is not an isolated German project, but rather an element of a larger European movement. The IPCEI-CIS program, which pools €1.2 billion in state aid from seven member states for cloud and edge computing technologies, provides the political and financial infrastructure for similar projects. Comparable consortium models exist in France with the Lucie model and at the pan-European level with the OpenGPT-X project. The commonality of these initiatives is structural: they combine public funding, academic capacity, and private infrastructure.

The context makes the difference clearer. Anyone expecting European-developed AI to compete with the multi-billion-dollar investments of OpenAI, Google, Anthropic, or the state-sponsored Chinese model ecosystem is asking the wrong question. The more relevant question is whether Europe is capable of building its own fully controllable layer of fundamental AI models that can serve as the basis for European application development—without complete dependence on non-European infrastructure, licensing terms, and geopolitics.

The EU AI Act, which is being phased in fully, adds a further legal dimension to this issue. For general-purpose models, it mandates transparency obligations that are structurally easier for fully open models with documented training data to fulfill than for proprietary black-box models. This is no coincidence: European regulation is partly designed to give European open-source approaches a comparative advantage over proprietary architectures. Soofi S fits this regulatory design perfectly.

An honest assessment of a first step

Soofi S is the first European open-source language model that not only boasts in press releases but also performs on par with international competitors in verifiable benchmarks – at least within the category of fully open models. This is no small feat. Its European predecessors played in a different league, and the gap was fundamental, not marginal.

At the same time, it would be intellectually dishonest to reinterpret this progress as an AI breakthrough, which it is not. A 30-billion-parameter model that lags behind Qwen3.5 and is still in the beta phase is a promising start, not an endpoint. The consortium's research quality is genuine. The architectural decisions are well-considered. The transparency is exemplary. But the gap to the global frontier remains significant, and it cannot be closed with just 20 million euros of public funding.

What sets Soofi S apart from all previous announcements of sovereign European AI is a single, crucial detail: the model actually exists, with published weights, documented training, and measurable results. This sounds obvious, but it still isn't in the European AI landscape. For those who treat data sovereignty, auditability, and GDPR compliance as genuine decision criteria—and not just compliance rhetoric—a new equation begins here.

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

Konrad Wolfenstein

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here [email protected]:or simply call me at +49 7348 4088 965. My email address is

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business