A German research consortium has released Soofi S, a 30-billion-parameter open language model that it says leads fully open rivals on combined English and German benchmarks while running with the inference cost of a much smaller system. The launch matters beyond one more model release because the team is making a direct case for European-built, German-focused foundation models that can compete on performance without relying on US hyperscaler APIs.

The release also arrives with an unusual caveat. According to reporting from The Decoder and the consortium’s own updated technical report cited there, the team found that test questions from the science benchmark GPQA had accidentally entered the training data through a dataset-splitting mistake. After the issue was identified by the community, the consortium removed GPQA from its evaluation, recalculated results across all compared models, and said the ranking did not change. That combination of ambitious benchmark claims and a public contamination correction makes Soofi S notable for both its technical design and the transparency test it now faces.

What Soofi S is trying to do differently

According to The Decoder, Soofi S 30B-A3B is a mixture-of-experts system with 31.6 billion total parameters but only about 3.2 billion active per token. The consortium reportedly adopted the hybrid design of Nvidia’s Nemotron 3 Nano, combining Mamba-2 layers with traditional attention layers rather than using a standard dense transformer throughout.

That architectural choice is central to the project’s pitch. In ordinary transformer models, long prompts increase the burden of the KV cache used for attention, which can drag down throughput as context windows expand or as more requests are served in parallel. The Soofi S design reportedly limits that effect because only a small subset of layers maintains that cache. The Decoder says the consortium measured nearly flat throughput from 4,000 to 256,000 tokens and reported that at 40,000 tokens with 32 parallel requests, Soofi S generated roughly eight times more tokens per second per GPU than dense models in the 14B to 24B range.

Those figures are vendor-reported measurements from the consortium’s materials, not independent third-party benchmarks. Still, if they hold up in deployment, they would make Soofi S relevant to teams building document-heavy applications, long-context copilots, and enterprise systems where latency and GPU utilization matter as much as raw benchmark scores.

The model was trained entirely on Deutsche Telekom infrastructure, specifically the company’s Industrial AI Cloud in Munich, according to The Decoder. That matters for a different reason: sovereign AI projects in Europe have often struggled to show both credible compute access and credible technical results. Soofi S is explicitly positioned as evidence that a regionally trained, openly released model can use local infrastructure and still contend in mainstream evaluations.

A German-first data strategy, without giving up English

The consortium’s biggest strategic bet appears to be data mix rather than architecture alone. The Decoder reports that German represented 7.2 percent of the first training phase and 15.3 percent of the second, compared with roughly 5 percent for all non-English languages combined in Nvidia’s Nemotron reference recipe.

According to that report, training covered about 27 trillion tokens across three phases: a broad first pass over web, code, math, and domain texts; a second pass over higher-quality material; and a shorter long-context phase with documents up to one million tokens. The German portion came from sources including HPLT, German Commons, FinePDFs, FineWiki, and the commercially licensed Genios archive of German news content. The team also used machine-translated and synthetic German text.

That data weighting appears to have paid off in the consortium’s own evaluation. The Decoder says Soofi S leads all fully open models on aggregate English and German scores in comparisons against 16 open rivals, including OLMo 3 32B and Apertus 70B. On Germany-specific knowledge, it reportedly tied Qwen3.5 35B-A3B on INCLUDE-DE. On programming tests, reported scores of 73.8 on HumanEval, 70.2 on MBPP, and 84.2 on German MBPP put it ahead of open-source peers in the consortium’s benchmark suite.

For European AI teams, that is the real commercial hook. A model that can handle bilingual enterprise workflows, code generation, and local factual or regulatory context without sacrificing English capability is easier to justify than a narrowly national model that performs well only on domestic tasks. If Soofi S proves robust in practice, it could be attractive for internal assistants, search and summarization tools, coding assistant products, and sector-specific systems serving German-speaking knowledge workers.

The benchmark issue changes how the launch will be judged

The strongest reason to treat the benchmark story carefully is that the contamination issue was not hypothetical. The Decoder reports that rephrased test questions from GPQA and GPQA Diamond ended up inside the QA-base dataset used during training, including English and machine-translated German versions.

According to the report, the cause was a labeling quirk on Hugging Face: GPQA had no separate training split, and its test content sat under the default label “train.” Because the pipeline selected material by split name, the benchmark leaked into the training corpus. A later audit reportedly found three additional benchmarks with the same pattern — TruthfulQA, BLiMP, and Inverse Scaling — although The Decoder says those were not part of the Soofi S evaluation.

The consortium’s response, as described by The Decoder, was to remove GPQA entirely from the evaluation and recompute the comparative rankings for all 16 models. The team also said it now cross-checks training data directly against benchmark questions instead of relying on split labels. Consortium participants cited by The Decoder argued that the public release of data made community detection possible and therefore validated the open approach.

That is a defensible point, but it does not erase the episode. For model buyers and developers, the incident reinforces a broader lesson: benchmark leadership is useful, but only if the data pipeline and evaluation process can survive outside scrutiny. The reported availability of roughly 152,000 individual results and a separate evaluation by partner Ellamind on held-out internal data helps the consortium’s case, but those are still closely tied to project participants rather than neutral outside labs.

Where Soofi S looks strong — and where it does not

The Decoder’s account suggests Soofi S has a clear strength profile rather than universal superiority. The model reportedly performs especially well on German and English aggregate tests and on code benchmarks, while also benefiting from efficient long-context serving behavior. Those traits align with practical enterprise workloads more than with one-off consumer chatbot demos.

But the same report notes weak spots. On German competition math, Soofi S reportedly scored 56 on Minerva MATH-DE, behind both Qwen3.5 35B-A3B and Gemma 3 27B. It also lagged on NaturalQuestions, which the authors reportedly linked to lower factual storage capacity from having only about 3 billion active parameters per token.

There was also a notable failure pattern in the RULER long-context test. According to The Decoder, when asked to extract frequently occurring words from long text, Soofi S’s hit rate dropped to around 3 percent beyond 32,000 tokens, while the comparable Nemotron model maintained 60 to 64 percent. The authors reportedly attributed that weakness to long-context training data that included long documents but not enough synthetic extraction-style tasks.

That distinction matters for AI builders. “Long context” is not one capability. A model can be efficient at processing long inputs yet still struggle with specific retrieval or extraction behaviors inside those inputs. Product teams considering Soofi S for enterprise AI should test their exact workflow shape: legal review, customer-support summarization, coding assistant use, research copilots, or multilingual knowledge retrieval may stress different parts of the system.

Evidence, claims, and what is independently verified

Most of the performance, efficiency, and training details in this story come from The Decoder’s report on the consortium’s technical materials. That means the key benchmark and throughput claims remain consortium-reported rather than independently replicated.

What appears strongest from the available evidence is the existence of the release itself, the architecture choice based on Nvidia Nemotron, the use of Deutsche Telekom infrastructure, the German-heavy training mix, and the consortium’s disclosure that GPQA test content leaked into training data and was later removed from evaluation.

What remains harder to verify from the current source set are the precise comparative performance margins over OLMo 3 32B, Apertus 70B, and Qwen3.5 35B-A3B, as well as the real-world serving advantages under production loads. Those claims may prove directionally accurate, but external reproductions will matter more than launch-day charts.

The licensing timeline also bears watching. The Decoder says instruct and reasoning variants are in beta and expected under a permissive license in the coming weeks, while a larger Soofi L model is already in training. Until those variants ship and developers can test them directly, the commercial usefulness of the family is still only partly visible.

What to watch next

First, watch whether independent evaluators reproduce the headline that Soofi S leads fully open models in German and English. Third-party tests on HumanEval, MBPP, RULER, and German-language suites will carry more weight now that the GPQA issue is public.

Second, watch deployment benchmarks rather than static model cards. If Soofi S really preserves throughput across long contexts better than dense peers, that could matter for enterprise AI budgets and for teams trying to serve many concurrent users on limited GPU capacity.

Third, watch adoption signals in German-speaking industries. A model trained on Deutsche Telekom infrastructure and optimized for German may appeal to customers with data residency, procurement, or sovereign AI requirements, but those buyers will also want evidence on reliability, safety, and support.

Finally, watch whether Soofi L and the upcoming instruct models keep the same transparency standard. The openness argument made by the consortium becomes stronger only if future releases continue to expose datasets, methods, and post-launch corrections to outside scrutiny.

Creati.ai perspective

Soofi S is interesting not because Europe launched another open model, but because it targets a more practical niche: bilingual enterprise performance plus long-context efficiency on local infrastructure. That is a more useful value proposition than abstract sovereignty alone. If the throughput claims survive replication, Soofi S could appeal to builders who need large-context systems without paying the serving costs usually associated with dense frontier-class models.

The contamination disclosure, however, is part of the story rather than a footnote. In today’s model market, transparency is becoming a competitive feature of its own. Soofi S may ultimately benefit from having corrected the mistake in public, but it will still need independent validation. For founders and enterprise teams, the right takeaway is not “ignore the model” or “trust the benchmarks,” but “test the open weights against your own workloads.” That is where Soofi S will either become a serious regional platform candidate or remain an impressive technical report.