German AI Consortium Releases Soofi S, an Open 30B Model That Tops Benchmarks in Both English and German

A German research consortium coordinated by the KI Bundesverband (German AI Association) has released Soofi S 30B-A3B, an open-source language model trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich. According to its pretraining report, the model achieves the highest scores on both English and German benchmarks among fully open models, surpassing previous leaders including OLMo 3 32B from the Allen Institute for AI and Apertus 70B from ETH Zurich and EPFL.

The model uses a resource-efficient hybrid architecture that activates only 3.2 of its 31.6 billion parameters per generated token, keeping processing speed nearly constant even with very long inputs. A deliberate focus on German training data—accounting for up to 15.3 percent of the training mix in its second phase—gives Soofi S a strong advantage in German-language tasks while maintaining competitive English performance. That combination addresses a recurring gap in European AI projects: matching international open-weight performance while documenting training methods, infrastructure, and data sources closely enough for external scrutiny.

Contamination Incident Disclosed in Updated Report (July 24, 2026)

On July 24, 2026, the consortium published version 3.0 of its pretraining report and documented a data contamination incident that the community discovered in the publicly disclosed training data after the initial release. The problem affected the QA-base dataset, which was intended to contain only rephrased training splits from 25 standard benchmarks—practice questions designed to teach the model test formats, not the actual test questions.

However, the dataset also included rephrased questions from the test set of the science benchmark GPQA, including the GPQA Diamond variant, in both English and machine-translated German. The report traces the root cause to a structural quirk of GPQA on Hugging Face: the benchmark has no separate training set, and all of its test material sits under the default label "train." Because the data pipeline selected content based on split names, the test set was inadvertently mixed in with the practice questions.

A full audit triggered by the discovery turned up three additional benchmarks with the same labeling pattern—TruthfulQA, BLiMP, and Inverse Scaling—though none of those are used in the Soofi-S evaluation. In response, the consortium dropped GPQA entirely from its evaluation suite and recalculated the overall results for all 16 compared models. According to the authors, the ranking order remained unchanged.

Notably, nobody caught the problem during training. The model did improve steadily on the affected GPQA tasks, climbing from 32.3 to 43.4 points, but that gain fell within the normal range and was indistinguishable from progress on other tests—there was no sudden spike to serve as a warning sign. Going forward, the team now automatically cross-checks all training data against test questions rather than relying on potentially misleading split labels.

Project participant Nicolas Flores Herr from Fraunhofer IAIS pointed to the incident as evidence that the open approach works, noting the community could only have identified the problem because the data was publicly available. The team has made roughly 152,000 individual results available for verification. A separate evaluation by consortium partner Ellamind, using its own deliberately withheld test data that the model was guaranteed not to have encountered during training, confirmed the results. The consortium states that the affected portion represents only a tiny fraction of the total corpus.

The fine-tuned instruct and reasoning variants are currently in beta testing and are expected to ship under a permissive license in the coming weeks. The larger Soofi L model is already in training.

Response to Overtraining Allegations (July 15, 2026)

After launch, critics argued that Soofi S was heavily "overtrained" by the standards of the classic Chinchilla scaling laws, which Google DeepMind published in 2022. Those laws describe how to balance model size and training data for a fixed compute budget, identifying a sweet spot of roughly 20 tokens per parameter.

Soofi S far exceeds that ratio. With approximately 27 trillion tokens and 30 billion parameters, the ratio lands at several hundred to one. Factor in only the 3.2 billion parameters active per token, and the ratio jumps to several thousand to one.

Michael Fromm, part of the project's technical leadership, pushed back on that criticism, arguing that those rules do not directly carry over to Mixture-of-Experts (MoE) architectures. "There's new research showing that the old scaling laws from dense models no longer apply to MoE architectures," Fromm said. He explained that individual experts benefit from seeing the same documents repeatedly, so duplicated data in a large, high-quality dataset poses less of a problem than it would with dense models. As a comparison, Fromm noted that Nvidia trained its own models on up to 25 trillion tokens.

A Lean Architecture Built for Long Contexts

Soofi S is a mixture-of-experts model containing 31.6 billion parameters in total but activating only about 3.2 billion per generated token. This places its compute cost closer to a 3B model than a conventional 30B model. The consortium adopted the architecture of Nvidia's Nemotron 3 Nano without modification—a hybrid design combining Mamba-2 layers with standard attention layers.

The critical difference from typical transformers lies in memory behavior. In conventional models, the KV cache that stores previous tokens for attention computation grows linearly with context length, becoming a bottleneck under long inputs and many parallel requests. Only 6 of Soofi S's 52 layers maintain such a cache at all.

The practical payoff is evident in generation throughput. At a context length of 40,000 tokens with 32 parallel requests, Soofi S generates roughly eight times more tokens per second per GPU than dense models in the 14 to 24 billion parameter range. While throughput drops significantly for conventional models as context grows, Soofi S stays nearly flat from 4,000 to 256,000 tokens. The only model showing similar behavior in the measurements is Alibaba's Qwen3.5 35B-A3B, which also uses a hybrid architecture.

A Training Mix Built Around German

The consortium processed approximately 27 trillion tokens across three phases. In the first phase, the model learned language fundamentals from roughly 20 trillion tokens drawn from a broad mix of web, code, math, and domain-specific texts. A second phase followed with about 6 trillion tokens from higher-quality sources designed to sharpen earlier-learned patterns. A shorter third phase extended the context window by training on very long documents of up to one million tokens.

The deliberate emphasis on German is central to the design. In the first phase, German constitutes 7.2 percent of the training mix; in the second phase, that share rises to 15.3 percent. By comparison, in Nvidia's Nemotron reference recipe, all non-English languages combined account for only about 5 percent. For German-speaking enterprises and public-sector users, that data balance is relevant because many technical, legal, administrative, and customer-service workflows depend on domain language that is not well represented by English-centric corpora.

For data sources, the consortium combined German web text from HPLT, the openly licensed German Commons corpus, German portions of FinePDFs and FineWiki, and the commercially licensed Genios corpus containing 193 million newspaper articles from 916 German publications. Machine-translated and synthetically generated German texts rounded out the mix.

Benchmark Results Across German, English, and Code

In evaluations against 16 other open models, Soofi S leads all fully open models on aggregate scores for both German and English, according to the report. Against every European sovereign baseline, the model comes out ahead on all German benchmarks in the suite, sometimes by double-digit margins.

On code benchmarks, Soofi S scores 73.8 percent on HumanEval, 70.2 on MBPP, and 84.2 on the German MBPP variant—the best results among open-source peers. On INCLUDE-DE, a test for Germany-specific regional knowledge, Soofi S ties for first place at 61.2 points with the larger Qwen3.5 35B-A3B. Compared to the Nemotron baseline, the German data recipe improved language proficiency by 15.1 points and the science test GPQA-Diamond by 9.6 points, without sacrificing English performance.

Soofi S underperforms in certain areas. On German competition math, it scores 56 points on Minerva MATH-DE, well behind Qwen3.5 35B-A3B (76.5) and Gemma 3 27B (65.6). It also lags on open factual retrieval in NaturalQuestions, likely related to having only 3 billion active parameters, which can store less world knowledge than a dense 27B model.

The RULER long-context test revealed a specific weakness: when the model must extract frequently occurring words from a long text, Soofi S's hit rate drops to around 3 percent beyond 32,000 tokens of context, while the comparable Nemotron model still manages 60 to 64 percent. The authors attribute this to the fact that their long-context training data contains many long documents but lacks synthetic data designed specifically for extraction tasks. On the remaining twelve RULER tasks, both models perform comparably.

Sovereign Infrastructure and Documented Openness

The training run took place between March and May on up to 512 Nvidia B200 GPUs at Deutsche Telekom's Industrial AI Cloud in Munich, totaling approximately 253,000 GPU-hours. According to the report, the facility runs entirely on renewable energy, is cooled with water from the Eisbach canal, and feeds waste heat into the surrounding Tucherpark neighborhood. Soofi S was among the first major training runs on this infrastructure.

The consortium behind Soofi comprises German research institutions and companies, coordinated by the German AI Association and funded by the German Federal Ministry for Economic Affairs and Energy as part of the European IPCEI-CIS program. Participants include the Fraunhofer Institutes IAIS and IIS, the German Research Center for Artificial Intelligence (DFKI), TU Darmstadt, the University of Würzburg, the L3S Research Center, the Berlin University of Applied Sciences, and AI companies Ellamind and Merantix Momentum. The project's goal is to build an open European AI model family that can run on sovereign infrastructure and be tested in industrial applications.

The researchers are releasing model weights along with selected intermediate checkpoints, the complete training and evaluation code, and a detailed data inventory listing raw token counts, epoch numbers, and effective contributions per source. Sources that were reviewed but excluded are also documented. According to the team, this means Soofi S meets the Open Source AI Definition 1.0 from the Open Source Initiative.

A stricter proposal for a European open-data definition, which would require every single training token to be freely distributable, is not met because of the 1.3 percent share of Genios data carrying a commercial license. The report states that approximately 99 percent of the training mix can be independently reconstructed. The exact license for the model's release has not yet been finalized.

As technical leader Michael Fromm writes, Soofi S positions itself between broadly multilingual European sovereignty projects like EuroLLM or Teuken, which cover many languages, and the highest-performing international open-weight models. According to the project website, the consortium is seeking industry partners for the next phase to test the model in applications involving technical documents, code generation, and agent-based systems.