Soofi S: What Germany’s Open LLM Delivers in Practice
In July 2026 a German consortium coordinated by the KI Bundesverband released its first language model: Soofi S, trained on Deutsche Telekom’s Industrial AI Cloud in Munich. Coverage was uniformly positive, and the benchmarks support it. For decision-makers at mid-size companies, though, a different question matters: can you actually work with it today? The answer is more nuanced than the headlines suggest.
What the model is technically
Soofi S is a mixture-of-experts model with 31.6 billion parameters, of which roughly 3.2 billion are active per token. That pushes compute requirements well below what a dense model of this size would demand. The architecture combines Mamba-2 layers with conventional attention layers, 52 in total, only six of which maintain a KV cache. Memory stays manageable even on long inputs. Nvidia’s Nemotron 3 Nano architecture served as the base.
All figures below come from the consortium’s technical report and the Hugging Face model card. Training covered roughly 27 trillion tokens across three phases, with German making up 7.2 percent in the first phase and 15.3 percent in the second. The sources are disclosed: HPLT, German Commons, FinePDFs, FineWiki and a licensed Genios corpus of 193 million newspaper articles. The run took place from 24 March to 13 May 2026 on up to 512 Nvidia B200 cards across 64 DGX nodes, consuming roughly 253,000 GPU hours.
Funding comes from Germany’s Federal Ministry for Economic Affairs and Energy under IPCEI-CIS. That is the same European sovereignty framework behind the Open Component Model. The logic is identical in both cases: create open building blocks that European vendors can combine independently of US platforms.
The licensing question: it is not open yet
This is where the model card matters more than the press releases. On Hugging Face, Soofi S sits under a closed-beta license, access is gated, and the card states plainly that this is a beta preview and a research artifact, explicitly not an open release. A permissive license without gating has been announced, but as of mid-August 2026 it is not in force.
In practice this means the model cannot simply be downloaded and tried out. Every variant on Hugging Face, including the quantised GGUF and FP8 derivatives, sits behind manual approval, and a fetch without access returns HTTP 401. The project site is unambiguous: the beta runs with partners already selected, and further applications for the test phase are not being accepted at present. Evaluating the model today is therefore not an option.
That distinction changes the assessment. A model whose terms of use can still change is a poor foundation for a product expected to run two years from now. Anyone who has worked through the difference between open weights and a genuine open-source license will recognise the pattern from our article on open weights versus open source: the word “open” gets used generously in AI, and the details rarely make it into the headline.
In fairness, the consortium communicates the beta status transparently on the model card. The shortening happens in the coverage. The base model also has no safety tuning, which is normal for a research artifact but means any production use requires your own post-training and your own guardrails.
Where the benchmarks hold and where they do not
The good numbers are real. Among fully open models, Soofi S posts the best aggregate scores for German and English, ahead of Olmo 3 32B and Apertus 70B, the latter at less than half the parameter count. The German aggregate lands at 85.3 points, ARC-Challenge-DE at 92.3 percent, and on code aggregates for both German and English Soofi S ranks first among 16 open base models. For a model with 3.2 billion active parameters those are notable results. Against European sovereign models such as Teuken-7B, EuroLLM or Salamandra it matches or beats every German benchmark in the suite, often by 10 to 30 points.
Three caveats belong to an honest reading. First, mathematics: on Minerva MATH-DE the model scores 56 points, while Qwen3.5 at comparable size reaches 76.5. Second, factual recall: on open-domain questions Soofi S trails denser models, which is structurally expected at 3.2 billion active parameters. Third, and most consequential in practice, long context.
The model advertises a context window of up to one million tokens. The report measures this with RULER against the Nemotron base and draws a careful distinction: across all 13 subtasks Soofi S trails by 6.8 points on average, and excluding the common word extraction subtask the gap narrows to 4.4 points, which the report itself calls roughly on par. On that one subtask, however, the model collapses beyond 32,000 tokens, from 64 points to 3. For RAG scenarios that matters, because extraction over long documents resembles this pattern. It is not a general context weakness, but it is worth measuring on your own data before settling an architecture.
There is also a transparency note in the report itself: a contamination incident affects four QA constituents, so GPQA and a withdrawn benchmark group are excluded from both aggregates. The consortium discloses the incident in a dedicated section. That the consortium documents the mistake speaks well of its practice, but it does qualify the comparability of individual numbers.
Who should be looking at it today
The model is interesting to companies whose core work involves German-language processing and where data sovereignty is a hard requirement. The order of work is unusual at the moment, though: with access closed, you start with your own requirements rather than with the model. Classification, summarisation and extraction from German technical texts are the task classes where benchmark scores most reliably translate into practical quality. The low active parameter count also makes it cheaper to run than dense models of comparable quality, which adds a new candidate to the hardware maths in our piece on running LLMs locally.
Three things argue against production use this quarter. The license is not final, the base model is unaligned, and the long-context weakness hits exactly the architecture most commonly built in enterprise settings.
A two-step approach makes sense. Right now you can sharpen the use case itself: which tasks should a model take on, what test data exists, what quality threshold has to be met? That evaluation harness is needed regardless and works with any model. Once the open release lands, announced for this year, the measurement against your existing stack follows. Anyone wanting to join as a pilot user can reach the consortium at contact@soofi.info, but should expect the partner group to be settled already.
Conclusion
Soofi S is a technically serious model and the strongest contribution so far to a European language-model ecosystem. It is not production-ready at this point, and the model card says so more plainly than much of the coverage about it. Anyone planning sovereign AI should evaluate it and follow its development, but should not build an architecture decision on it before the licensing question is settled. For assessing which model fits which use case, and for building the infrastructure around it, we help through our AI consulting for the mid-market.
Sources
All technical details and benchmark figures in this article come from the Soofi consortium’s technical report, the Hugging Face model card and the project site soofi.info. Access status was last verified directly against the Hugging Face API on 18 August 2026.
Frequently Asked Questions
What is Soofi S?
Soofi S is an open language model for German and English, built by a German consortium coordinated by the KI Bundesverband and funded by the Federal Ministry for Economic Affairs and Energy. It has 31.6 billion parameters, 3.2 billion of them active per token, and was trained on Deutsche Telekom’s Industrial AI Cloud in Munich.
Is Soofi S really open source?
Not yet. The Hugging Face model card labels it a beta preview and research artifact, access is gated, and the license is a closed beta. A permissive license without gating has been announced but was not in force as of mid-August 2026.
Can Soofi S be used in production?
Not currently, and testing it is not possible either: every variant sits behind manual approval, and the project is not accepting applications for the test phase at present. Beyond that, the open licensing question, the missing safety tuning and the weak performance on very long contexts argue against production use.
How does Soofi S compare to other European models?
On aggregate scores for German and English, Soofi S ranks ahead of Olmo 3 32B and Apertus 70B, despite having less than half the parameters of the latter. Against European sovereign models such as Teuken-7B, EuroLLM and Salamandra the report has it matching or beating every German benchmark in the comparison suite, often by 10 to 30 points.
