Soofi S: A Sovereign AI Model from Germany, and Why the Details Matter

Most enterprises run their AI on models they cannot inspect, hosted on infrastructure they do not control, trained on data nobody will show them. For a demo, that is fine. For a regulated process that touches customer data, contracts, or safety documentation, it is a standing liability you have quietly agreed not to look at. In July 2026 a German research consortium published something that lets you look: Soofi S, a 30-billion-parameter foundation model whose weights, training recipe, and per-source data accounting are all on the table.

I read the paper so you do not have to. Here is what Soofi S actually is, why the "made in Germany" part is more than a flag on the box, and where it could earn its place in a real business stack.

What Soofi S is

Soofi S was introduced in the paper "A Sovereign, Open-Source Foundation Model for German and English" (arXiv:2607.09424), published by The Soofi-Team on 10 July 2026. The name is a programme as much as a label: SOOFI stands for Sovereign Open Source Foundation Models.

It was not built by a single lab. The consortium is coordinated by the KI Bundesverband (the German AI Association) and brings together Fraunhofer IAIS and IIS, the DFKI, TU Darmstadt, the University of Würzburg, the L3S Research Center in Hannover, and companies such as ellamind and Merantix Momentum. It was funded by the German Federal Ministry for Economic Affairs and Energy under the European IPCEI-CIS programme, and trained end to end on the German Industrial AI Cloud, a sovereign HPC-scale facility operated by Deutsche Telekom in Munich.

The model itself is Soofi S 30B-A3B: 31.6 billion parameters in total, of which only about 3.2 billion are active for any given token. It is deliberately a German-English specialist, described by its authors as "a German-English champion rather than a thinly spread multilingual model."

The architecture, and why it is cheap to run

Soofi S is a Mixture-of-Experts (MoE) hybrid Mamba-Transformer. Two ideas do the heavy lifting, and both matter for your bill.

Translated into business terms: long contracts, entire codebases, retrieval over large document sets, and many concurrent users, without the runaway cost curve of a dense frontier model. For the kind of document-heavy, high-concurrency workloads that SAP-centric enterprises actually run, that shape is the whole point.

Radical transparency about the training data

Soofi S was pretrained on roughly 27 trillion tokens across a three-phase curriculum, with German deliberately up-weighted: from 7.2 percent of the mix in the first phase to 15.3 percent in the second. The sources are named, including German web text from HPLT, the German Commons corpus, FineWiki and FinePDFs, code and mathematics collections, and the commercially licensed Genios corpus of 193 million German newspaper articles.

The headline is the openness. The team released it, in their own words, "radically open: not weights alone, but the full set of artifacts required to audit every stage of training." Weights, intermediate checkpoints, hyperparameters, full per-source data accounting, and the training and evaluation code are all published. About 99 percent of the data mixture can be independently reconstructed, and the paper even lists the sources it considered and *excluded*, so the recipe can be audited and rebuilt end to end. The one exception, the Genios newspaper corpus at 1.3 percent of the first phase, is documented with aggregate statistics rather than raw redistribution, because its licence is commercial. The result meets the Open Source Initiative's Open Source AI Definition 1.0. The weights are released openly under permissive terms; the exact final licence was still being finalised at publication, so treat it as open-but-verify rather than a specific named licence.

How good is it, really

The authors are careful with their claims, and so am I. In their own framing, Soofi S "matches dense 14 to 27B models on aggregate English and German benchmarks while achieving the best code aggregates in both languages among 17 open base models, and outperforms every European sovereign baseline in our comparison, including ones far larger in active parameters. Among fully open models, Soofi S obtains the highest English and German evaluation scores, ahead of Olmo 3 32B and Apertus 70B."

Some concrete numbers back that up: a German aggregate of 79.1, 73.8 on HumanEval and 84.2 on the German MBPP for code, 86.1 on GSM8K for grade-school maths, and 92.3 on the German ARC-Challenge. It does not beat the strongest closed or semi-open frontier models on every aggregate. That is not the claim, and it is not the point. The point is that this is the strongest result you can get from a model where you are allowed to see everything, in the two languages a German business most cares about.

Why "from Germany" is more than a flag

It is easy to be cynical about the sovereignty framing. It is also a mistake, because the dependency is real and mostly invisible until it bites.

Most enterprises today route their AI through American or Chinese models. That is a strategic and legal dependency, not just a technical choice. Pricing, availability, terms of service, model deprecations, and export politics all sit outside your control. When your core processes depend on an API in another jurisdiction, someone else owns your roadmap. As Jörg Bienert of the KI Bundesverband put it when the project was announced: "Whoever controls the foundation models controls a central part of future digital value creation."

Soofi S changes the geometry. It runs on sovereign infrastructure in Munich and can be self-hosted on-premise, which means prompts that carry customer data, contract text, or engineering documentation never have to leave EU jurisdiction or your own datacenter. For a manufacturer's technical documentation or a bank's regulatory filings, that is frequently a hard requirement, not a nice-to-have. And under the EU AI Act, general-purpose AI models carry transparency obligations, including a summary of the content used for training. A model whose data is already documented source by source hands you most of that paperwork; a closed model leaves you dependent on whatever a foreign vendor chooses to disclose.

And Soofi S does not arrive into an empty field, which I am genuinely glad about. Aleph Alpha in Heidelberg has been at this for years: the Luminous family (up to 70 billion parameters, trained on a curated corpus of English, German, French, Italian and Spanish) and, more recently, Pharia-1-LLM, a 7-billion-parameter model released in 2024 and trained, in the company's words, on "carefully curated data in accordance with applicable EU and national regulations, including copyright and data protection law." Different bets, the same instinct. I am happy to see more capable models coming out of the domestic market, and equally honest about the gap: against the American and Chinese frontier, we in Germany still have real ground to make up. Soofi S is a strong step, not a finish line.

Why knowing the training data is a business feature

The "you can see the data" point sounds academic. It is not. It converts several vague risks into inspectable facts.

None of this hands you a rubber stamp. Putting a model into a regulated process still takes verification steps and a lot of testing, and no amount of openness removes that. But knowing how a model was trained makes that trust far easier to earn, and curation is increasingly the enterprise path to it. IBM made the same bet with its Granite family: curated and documented so deliberately that in 2025 it became the first open-model developer to have the AI management system behind its models certified to ISO/IEC 42001, the international standard for managing AI. A model earns enterprise trust not by winning a leaderboard but by having a provenance and a process you can point to. AI is enterprise-ready the moment it is safe to use. Everything else is research, or playing around.

The business benefit, concretely

Strip away the politics and a practical case remains. The low active-parameter count and the near-constant long-context cache make inference cheap at scale, which is exactly what retrieval, agents, and bulk document processing need. The permissive licence means you can build a product on it, not just experiment. And the consortium is explicit that this is a starting point, not a finished product: the base model is open now, while specialised variants for dialogue, reasoning, and agentic use are being tested with industry partners on technical and regulatory documents, code generation, and agent-based systems. If your business runs on German-language technical content, that is an invitation to shape the tooling rather than rent it. The weights are already on Hugging Face, and I have asked to join the test programme myself, to verify Soofi S against my own use cases.

What I would watch

I would not oversell it. Three honest caveats.

Conclusion

The interesting thing about Soofi S is not that it is German. It is that being German, in this case, comes bundled with something rarer: you are allowed to see how it was made. In a market where the strongest models are also the most opaque, a model you can audit, self-host, and legally adapt is a different kind of asset, one whose provenance you can actually show a compliance officer. It will not top every leaderboard. But for a German or European business that has to know what is in its stack, that is often exactly the right trade.

If you are weighing a sovereign or open model for an SAP or enterprise workload, or deciding what to self-host versus buy, that is precisely the kind of decision I help with at AI & SAP Consulting Paul Oesterwitz.

Sources