Soofi S: Is Germany's Sovereign LLM Ready for Business?

Soofi S is a new German-English foundation model built by a German research consortium, trained from scratch on Deutsche Telekom's Industrial AI Cloud in Munich. It has 30 billion total parameters, activates roughly 3 to 3.5 billion per token, and was pretrained on about 27 trillion tokens. The consortium's own evaluation puts it ahead of OLMo 3 32B and Apertus 70B on aggregate English and German scores among the fully open models it tested.

That makes Soofi S technically important. It does not yet make it a production-ready European replacement for Claude, GPT, or every Chinese open-weight model. As of 16 July 2026, the public Hugging Face artifacts are labelled preview checkpoints, downloads require sharing contact information, the model card lists a custom license whose full text is still a TODO, and several safety, privacy, evaluation, and dataset fields remain unfinished.

The commercial answer is therefore more useful than the launch headline: Soofi S belongs on the pilot shortlist for German-language industrial AI, but not yet on an unqualified production shortlist. This review separates what the project has demonstrated from what a buyer still needs to verify.

What is Soofi S?

Soofi stands for Sovereign Open Source Foundation Models. The project is coordinated by the German AI Association and brings together Fraunhofer IAIS and IIS, DFKI, TU Darmstadt, the University of Würzburg, Leibniz University Hannover, Berlin University of Applied Sciences, ellamind, and Merantix Momentum. It is funded by Germany's Federal Ministry for Economic Affairs and Energy.

The first model, Soofi S 30B-A3B, is a sparse Mixture-of-Experts model. It combines Mamba-2 state-space layers with attention layers and activates only a fraction of its 30 billion parameters for each token. This matters because active parameters drive much of the compute per generated token, while total parameters still determine weight memory and storage.

The architecture is aimed at long documents and concurrent inference. Only a small number of layers maintain a conventional attention cache, so the cache grows much more slowly than it does in a dense Transformer. In the Soofi S pretraining report, the authors measured up to 9.2 times the aggregate decode throughput of dense 14B to 24B models at 40,000 tokens and batch size 32 on a single NVIDIA B200. That is a specific serving test, not a promise that every deployment will be nine times faster.

The training provenance is equally central to the project. Soofi S was trained end to end on up to 512 B200 GPUs in Deutsche Telekom's Munich facility. Deutsche Telekom describes the Industrial AI Cloud as infrastructure for European industry, research, start-ups, and the public sector. For a buyer, that proves where the pretraining happened. It does not automatically prove where your prompts, retrieval data, logs, or backups will be processed after deployment.

Is Soofi S really Europe's first fully European LLM?

No. "Europe's first LLM" is not an accurate description. Europe already had several models developed and trained by European teams. The German OpenGPT-X project published Teuken-7B, covering all 24 official EU languages, in 2024. EuroLLM-22B was trained from scratch for 24 EU languages and 11 additional languages. Switzerland's Apertus 70B released weights, data reconstruction code, intermediate checkpoints, and training recipes under Apache 2.0.

The defensible wording is narrower: Soofi S is the first model in the Soofi family, a German-developed and German-trained sovereign foundation model with unusually ambitious openness and strong German-English performance. That is still significant. It just is not a historical first for European language models.

It is also not "fully European" if that phrase means every dependency came from Europe. The model was trained in Germany and governed by a German consortium, but the training stack and hardware include NVIDIA technology. Sovereignty is not autarky. The practical goal is control over the model artifacts, infrastructure, data flows, and right to operate without one non-European model vendor deciding access or price.

Is Soofi S fully open-source or only open-weight?

The project's openness target goes well beyond weights. The pretraining report says the final release will include model weights, selected intermediate checkpoints, per-source data accounting, hyperparameters, and training and evaluation code. Where source licenses permit, the consortium plans to release data-construction artifacts. Commercially licensed sources are documented through mixture accounting rather than redistributed as raw text.

That direction aligns more closely with the Open Source AI Definition than a normal open-weight release does. The OSI definition expects the parameters, the code used to train and run the system, and sufficiently detailed information about the training data to let a skilled team recreate an equivalent system.

But buyers procure what exists, not what a paper promises to publish. The current Soofi S Instruct model card says:

• the checkpoint is a preview and its weights and metadata may change;
• the model is for research and development;
• the license is "Other," with the full license text still to be added;
• downloads are gated behind contact-information sharing;
• the final release date, cumulative compute, context limit, benchmark section, and several privacy and safety fields remain TODOs.

So the precise answer on 16 July 2026 is: Soofi S is designed to become a fully open model, but the current public preview is not yet complete enough for a company to approve commercial production use on the strength of the "fully open" label alone. Legal and security teams should wait for the final license, artifact inventory, model card, and immutable release hashes.

How strong is Soofi S on German, English, coding, and reasoning?

The benchmark headline is real within a defined category. In the consortium's evaluation of 17 open base models, Soofi S scored highest among the models classified as fully open on the authors' English and German aggregates. It ranked ahead of OLMo 3 32B and Apertus 70B, and it led the tested fully open group on code aggregates in both languages.

This is where the LinkedIn claim needs surgery. Soofi S does not top a universal ranking of all open models. It tops the fully open subset used in the Soofi team's aggregate evaluation. In the broader open-weight comparison, Qwen3.5 35B-A3B scored higher on both the English and German aggregates. Soofi S also trails Qwen3.5 on German competition mathematics, while performing particularly well on several coding measures and German regional knowledge.

The authors also report a long-context weakness. The architecture keeps serving throughput high as context grows, but a RULER extraction task involving frequent words deteriorated sharply beyond 32,000 tokens. Fast long-context generation and accurate long-context retrieval are different properties. A procurement test needs to measure both.

None of these base-model results proves that Soofi S matches the best Claude, GPT, Gemini, GLM, or Qwen production assistant on an end-to-end business workflow. Base-model benchmarks isolate pretraining quality. A production assistant also depends on post-training, tool use, retrieval, guardrails, serving reliability, and support.

Where could Soofi S create commercial value?

The model is most interesting when German-language quality, infrastructure control, and transparency matter more than winning every frontier benchmark.

• Technical and regulatory document workflows. German manuals, standards, policies, tenders, and internal knowledge bases are a natural evaluation target. Use retrieval and citation checks rather than expecting the model to contain current documents.
• Industrial coding assistants. The reported code scores justify a pilot for code explanation, test generation, migration support, and internal developer tooling. Repository-level acceptance tests still decide whether it ships.
• Agentic back-office workflows. The sparse architecture may suit concurrent agents processing long documents, provided tool-use accuracy, permissions, and failure handling pass a real eval.
• Domain adaptation. A transparent base model can be continued-pretrained or fine-tuned for an industry's language without depending on a proprietary vendor's roadmap.
• Public-sector and regulated pilots. European training provenance and the option to operate on controlled infrastructure can simplify part of the sovereignty discussion. It does not remove procurement, security, data protection, or model-risk work.

For these use cases, model choice is only one layer. Wavect's AI enablement work covers the surrounding system: use-case selection, data access, retrieval, evaluation, routing, deployment, and knowledge transfer. The proof pattern is visible in our Twinsoft AI case study, where the valuable part was the production system around the model, not a leaderboard position.

When should a company not choose Soofi S?

Do not choose it today when any of these conditions is true:

• You need a signed commercial license now. The preview model card does not yet provide one you can clear.
• You need a vendor SLA and managed support. The current public release is a research preview, not a managed enterprise service.
• Your workload is mainly multimodal or highly multilingual. Soofi S focuses on German and English. Qwen3.5, Apertus, or EuroLLM may fit broader language or modality requirements better.
• You need the strongest available abstract reasoning. The model's strongest evidence is German, coding, and serving efficiency, not universal frontier leadership.
• Your volume is low and your data can use a managed API. Running a 30B-total-parameter model can cost more in engineering and idle infrastructure than the API bill it replaces. Use our local-model versus API break-even calculator for that decision.
• You cannot build and maintain an eval harness. A sovereign model without measured task quality is still an unmanaged production risk.

What should buyers verify before a Soofi S pilot?

1. Freeze the exact artifact. Record the repository, revision, weight hashes, model card, custom code, tokenizer, and quantization. Preview artifacts can change.
2. Clear the final license. Confirm commercial use, modification, redistribution, output terms, and obligations for derivatives. Do not infer rights from the project name.
3. Choose the right variant. The base model is not a chatbot. Test the Instruct preview for general assistant work and evaluate reasoning variants separately rather than transferring base-model scores to them.
4. Build a use-case eval. Include at least 50 to 100 representative German and English tasks, hard negatives, unacceptable outputs, citation checks, tool-call accuracy, and human scoring.
5. Benchmark the target serving stack. Measure prefill, time to first token, decode throughput, concurrency, total GPU memory, failure recovery, and quantization loss on the hardware you will actually operate.
6. Map the entire data path. Document embeddings, vector storage, prompt logs, observability, backups, support access, and fallback APIs. A European model can still sit inside a non-sovereign system.
7. Price the successful task. Compare Soofi S with at least one European fully open alternative, one permissively licensed open-weight model, and one managed API. Include engineering, redundancy, monitoring, and eval maintenance.

Our technology selection guide explains the broader principle: choose against the constraint that is expensive to reverse. For Soofi S, that is currently license and operational maturity first, task quality second, then throughput and cost.

Soofi S vs Apertus, OLMo 3, and Qwen3.5

This table does not replace our broader open-weight LLM comparison. That page owns the market-wide choice among DeepSeek, Qwen, Kimi, GLM, and Llama. This page owns one narrower decision: whether Soofi S is ready for a European organization's pilot and procurement process.