German consortium releases Soofi S, a 30B model built for German and English

The German Soofi consortium has released Soofi S 30B-A3B, a bilingual foundation model that targets industrial applications such as technical documents, code generation and AI agents. The project trained the model on Deutsche Telekom’s Industrial AI Cloud in Munich and published its pretraining report, training code and evaluation tools.

Soofi S occupies a specific corner of the European model market. EuroLLM, Teuken and other projects spread capacity across many languages. Soofi focuses on German and English, with a training recipe built around German-language data and a serving design aimed at long documents.

The consortium says Soofi S leads fully open models on its combined English and German evaluations, ahead of OLMo 3 32B and Apertus 70B. The claim comes from the consortium’s own comparison of 17 open base models, so buyers should treat it as a reported result rather than an independent ranking.

A sparse model for long inputs

Soofi S contains 31.6 billion parameters, but the routing system activates about 3.2 billion parameters for each token. That mixture-of-experts design reduces the computation required for each generation step.

The model follows Nvidia’s Nemotron 3 Nano architecture. The design combines Mamba-2 layers, mixture-of-experts layers and grouped-query attention. Mamba-2 carries much of the sequence processing through a fixed-size recurrent state. Six of the model’s 52 layers use an attention cache, while conventional transformer models store a key-value cache across every layer.

That memory pattern affects long-context serving. In the consortium’s tests, Soofi S generated about eight times more tokens per second per GPU than dense models with 14 billion to 24 billion parameters at a 40,000-token context and 32 parallel requests. Soofi S kept a near-flat throughput curve from 4,000 to 256,000 tokens. Qwen3.5 35B-A3B showed a similar pattern because it uses a hybrid architecture.

The result matters for companies that process large manuals, contracts, code repositories or collections of internal records. A model can accept a long input and still serve several requests without letting cache memory dominate the workload. That advantage depends on the serving stack, hardware, quantization and request pattern, so the consortium’s measurements do not predict every production deployment.

German data drives the training recipe

The consortium trained Soofi S on about 27 trillion tokens across three phases. The first phase supplied broad material from web pages, code, mathematics and specialist domains. The second phase raised the share of higher-quality sources. The final phase used documents with up to 1 million tokens to extend the context window.

German data made up 7.2% of the first phase and 15.3% of the second. The corpus combined German web text from HPLT, German Commons, German subsets of FinePDFs and FineWiki, and the Genios archive of newspaper articles. Machine-translated and synthetic German text filled out the mixture.

Genios adds a disclosure issue. The consortium says the commercial corpus accounts for 1.3% of effective tokens in the first phase. The team documents that material through aggregate statistics, but it cannot redistribute the source text. That limitation prevents Soofi S from meeting stricter definitions that require every training token to remain freely available.

The project says its documentation meets the Open Source AI Definition 1.0 because the release includes weights, selected checkpoints, training details, data accounting and code. The Hugging Face model card lists the current base model as a preview with a custom “Other” license and says the full license text remains unfinished. The project website says the consortium has not opened general public use yet and is seeking industrial testing partners.

Strong results with clear gaps

Soofi S scored 73.8% on HumanEval, 70.2% on MBPP and 84.2 on the German MBPP test, according to the report. The model tied Qwen3.5 35B-A3B on INCLUDE-DE at 61.2 points, a benchmark for German regional knowledge.

The German-focused recipe produced the largest gains in German language tests. Soofi S placed first among the comparison models across the report’s German benchmark groups, including GLP-DE, ARC-Challenge-DE and MBPP-DE.

The model does not win every task. Soofi S scored 56 on Minerva MATH-DE, behind Qwen3.5 35B-A3B at 76.5 and Gemma 3 27B at 65.6. It also trailed on NaturalQuestions, where a model with 3.2 billion active parameters may store less factual information than a dense model with a larger active network.

The RULER long-context test exposed another weakness. Soofi S’s hit rate for extracting frequent words fell to about 3% beyond 32,000 tokens. Nemotron 3 Nano reached 60% to 64% on the same task. The researchers linked the gap to long-document training without synthetic extraction examples. Soofi S matched Nemotron on the other 12 RULER tasks in the report.

A sovereign training run

The consortium ran the training from March through May on as many as 512 Nvidia B200 GPUs. The run consumed about 253,000 GPU-hours at Deutsche Telekom’s Munich facility. The project says the site uses renewable energy, draws cooling water from the Eisbach canal and sends waste heat to the nearby Tucherpark district.

The KI Bundesverband coordinates the project. Fraunhofer IAIS, Fraunhofer IIS, the German Research Center for Artificial Intelligence, TU Darmstadt, the University of Würzburg, the Berlin University of Applied Sciences, the L3S Research Center, ellamind and Merantix Momentum participate in the consortium. Germany’s Federal Ministry for Economic Affairs and Energy funds the work through the European IPCEI-CIS program. The available project materials do not disclose a total funding amount or investor ownership structure.

Soofi’s market position depends on two promises: German-language performance and control over the training environment. The consortium trained the model on German infrastructure, documented its data sources and published training code plus evaluation code. Those materials give companies more visibility into the model’s origins than a weights-only release.

The official Soofi site lists technical and regulatory documents, code generation and agent systems as target applications. The project is asking companies to test those use cases before it defines a broader deployment path.

The overtraining dispute

Critics questioned the training budget after the launch. Soofi S saw about 27 trillion tokens against a model size near 30 billion parameters, far above the token-to-parameter ratio associated with the Chinchilla scaling laws for dense models.

Technical lead Michael Fromm rejected a direct comparison. He said research on mixture-of-experts models suggests that dense-model scaling rules do not transfer without adjustment. Each expert sees a narrower slice of the computation, and repeated examples in a large, high-quality corpus may help experts specialize.

The report does not settle that debate. Soofi S’s benchmark results show that the training approach produced a capable bilingual base model, while its weaker math and retrieval scores show the cost of allocating capacity toward German depth, code and long-context serving. Independent evaluations will decide whether the approach transfers to production workloads.

The model repository gives developers access to the project’s release materials and related checkpoints. The base model has no instruction tuning or safety alignment, so developers must add post-training, testing and safeguards before they place it in an end-user product.