A German research consortium has released Soofi S 30B-A3B, an open large language model positioned as both a performance play and a sovereignty play for Europe’s AI stack. According to reporting by The Decoder citing the model’s pretraining report, the system was trained entirely on Deutsche Telekom infrastructure in Munich and is designed to deliver strong German-language performance without giving up competitiveness in English and coding tasks.

The launch matters because Europe’s open-model efforts have often split into two camps: multilingual public-interest projects that emphasize coverage and governance, and global open-weight systems that win on benchmarks but are usually shaped by US or Chinese labs. Soofi S tries to bridge that gap. The consortium says the model now leads fully open peers on aggregate German and English benchmarks, while also using an architecture intended to keep inference efficient at long context lengths.

That combination could make the release relevant beyond research circles. For enterprise buyers, Soofi S is being framed as a model that can run on regional infrastructure with transparent training documentation. For builders, it is another sign that open-model competition is shifting from parameter counts alone toward a mix of data curation, inference efficiency, and jurisdictional control.

What Soofi S is and who built it

According to The Decoder, Soofi S is a mixture-of-experts model with 31.6 billion total parameters, but only about 3.2 billion are active for each generated token. The consortium calls the model Soofi S 30B-A3B, reflecting that sparse activation pattern. In practical terms, that means its compute profile can look closer to a much smaller model than a dense 30B system, at least for token generation.

The project was coordinated by the KI Bundesverband, or German AI Association, and funded through Germany’s Federal Ministry for Economic Affairs and Energy under the European IPCEI-CIS program, according to the report cited by The Decoder. Participating institutions include Fraunhofer IAIS, Fraunhofer IIS, DFKI, TU Darmstadt, the University of Würzburg, the L3S Research Center, Berlin University of Applied Sciences, Ellamind, and Merantix Momentum.

The infrastructure angle is central to the story. The model was trained on Deutsche Telekom’s Industrial AI Cloud in Munich, with the training run taking place from March to May on as many as 512 Nvidia B200 GPUs, according to The Decoder’s account of the technical report. The publication says the run consumed about 253,000 GPU-hours. Deutsche Telekom’s facility is described in the report as operating on renewable energy, using canal water cooling, and feeding waste heat into the surrounding district, though those sustainability details come from the project’s own materials rather than independent verification.

Architecture and training choices are the real product story

The technical design helps explain why the consortium is emphasizing throughput as much as raw benchmark performance. The report says Soofi S adopts Nvidia’s Nemotron 3 Nano architecture without modification, using a hybrid design that mixes Mamba-2 layers with conventional attention layers.

That matters because long-context inference has become one of the most expensive and operationally awkward parts of serving frontier-style models. Standard transformer models rely on a key-value cache that grows with context length. As inputs get longer and concurrency increases, memory bandwidth and cache movement become a bottleneck. According to The Decoder, only six of Soofi S’s 52 layers maintain that cache, which changes how the model scales under long prompts.

The consortium claims that at 40,000 tokens of context and 32 parallel requests, Soofi S produces about eight times more tokens per second per GPU than dense models in the 14B to 24B range. The report also says throughput stays nearly flat from 4,000 to 256,000 tokens, with Alibaba’s Qwen3.5 35B-A3B identified as the only other measured model showing similar behavior. Those are important claims for anyone building document-heavy workflows, retrieval pipelines, or agent systems that carry large working contexts, but they remain consortium-reported measurements rather than independent third-party benchmarking.

The training recipe is equally notable. The consortium says it processed roughly 27 trillion tokens across three phases: a broad first stage of about 20 trillion tokens spanning web, code, math, and domain-specific text; a second phase of roughly 6 trillion higher-quality tokens; and a shorter final phase designed to extend the context window with documents up to one million tokens.

What differentiates Soofi S from many open peers is its heavier German weighting. The Decoder reports that German accounted for 7.2 percent of the first phase and 15.3 percent of the second, compared with around 5 percent for all non-English languages combined in Nvidia’s Nemotron reference recipe. Training data reportedly included HPLT, German Commons, FinePDFs, FineWiki, and the commercial Genios corpus, plus machine-translated and synthetic German text.

Benchmark wins come with caveats

The headline claim is straightforward: the consortium says Soofi S leads all fully open models on aggregate English and German benchmark scores. The models it reportedly beat include OLMo 3 32B from the Allen Institute for AI and Apertus 70B from ETH Zurich and EPFL. The report also says it outperformed European sovereign baselines across all German benchmarks in the comparison suite.

On coding tasks, The Decoder says Soofi S posted 73.8 percent on HumanEval, 70.2 on MBPP, and 84.2 on a German MBPP variant. On INCLUDE-DE, a benchmark for Germany-specific regional knowledge, it reportedly tied Qwen3.5 35B-A3B at 61.2 points. The report further claims that, relative to the Nemotron baseline, the German-focused data mix improved language proficiency by 15.1 points and GPQA-Diamond by 9.6 points without harming English performance.

Those are strong results if they hold up under wider testing, but they should be read carefully. The available evidence in this news cluster comes from The Decoder’s reporting on the project’s own pretraining report and website materials. There is no independent benchmark audit in the source set, and no external reproduction of the throughput or quality claims.

The model also shows clear weaknesses. The Decoder reports that Soofi S underperformed on German competition math, scoring 56 on Minerva MATH-DE, behind both Qwen3.5 35B-A3B and Gemma 3 27B. It also lagged on NaturalQuestions, which the authors reportedly link to the lower active-parameter count versus dense models that may retain more factual knowledge.

Long-context behavior is not uniformly strong either. On the RULER test, the model apparently struggled on extraction tasks involving frequently repeated words in long documents. Beyond 32,000 tokens, its hit rate dropped to around 3 percent, while the comparable Nemotron model still managed roughly 60 to 64 percent, according to The Decoder’s summary of the report. The authors attribute that gap to a lack of synthetic extraction-focused data in long-context training.

Open model, but not open on every definition

Soofi S is being released with model weights, selected intermediate checkpoints, training and evaluation code, and a detailed data inventory, according to The Decoder. That degree of disclosure is significant in a market where “open” often means only downloadable weights with limited provenance information.

The consortium says this package satisfies the Open Source AI Definition 1.0 from the Open Source Initiative. At the same time, the report acknowledges it would not meet a stricter proposed European open-data standard because about 1.3 percent of the training mix came from the commercially licensed Genios corpus. The team says about 99 percent of the dataset can be independently reconstructed, but the exact release license had not been finalized at the time of the report.

That nuance matters for enterprise procurement and downstream product teams. For some users, open weights plus code and documentation are enough. For others, especially public-sector buyers and companies with strict redistribution or audit needs, the unresolved license details and inclusion of commercially licensed data may remain a meaningful constraint.

Why this matters for builders and enterprise buyers

For AI builders, the main lesson from Soofi S is that regional specialization can still move the needle. Instead of trying to be maximally multilingual, the consortium appears to have concentrated on German-language quality while preserving broad English utility. That is a practical strategy for teams serving regulated industries, technical documentation, customer support, and internal knowledge workflows where local-language performance is more valuable than universal coverage.

For enterprise AI deployments, the combination of open weights, German-tuned performance, and training on Deutsche Telekom infrastructure gives the model a clear positioning around data residency and sovereign infrastructure. That does not automatically solve compliance or risk concerns, but it does create a more concrete option for organizations that want alternatives to closed US platforms or opaque open-weight releases.

The architecture may prove just as important as the benchmarks. If Soofi S can sustain throughput at very long context lengths in real production settings, it could be useful for document review, coding assistant workflows, and AI agents that need to maintain large working memory. But those use cases will depend on whether the model’s extraction weakness on RULER shows up in customer workloads. A fast long-context model is less compelling if it misses key retrieval or extraction tasks inside those long documents.

Evidence and claims

Most of the substantive evidence in this story comes from The Decoder’s reporting on the consortium’s pretraining report, project website, and lead-author commentary. As a result, the strongest claims in this article, including benchmark leadership, throughput gains, training efficiency, and open-source compliance, should be treated as project-reported rather than independently verified.

What appears solid from the available sources is the existence of the release, the model’s basic architecture, the involvement of the KI Bundesverband-led consortium, the use of Deutsche Telekom infrastructure, and the publication of weights, code, and documentation. The comparative performance and operational advantages are more provisional until external users test Soofi S against OLMo 3 32B, Apertus 70B, Qwen3.5 35B-A3B, Gemma 3 27B, and Nvidia Nemotron in reproducible settings.

What to watch next

The first signal to watch is licensing. The exact release terms will shape whether Soofi S is mainly a research artifact or a practical base model for commercial product teams.

Second, look for independent benchmarking and inference tests. External comparisons on HumanEval, MBPP, long-context workloads, and German enterprise document tasks will matter more than leaderboard positions reported by the project itself.

Third, the consortium says it is seeking industry partners for the next phase in technical documents, code generation, and agent-based systems. Real deployment references in those areas would be a stronger proof point than benchmark wins alone.

Finally, watch whether Soofi S becomes a template for other European sovereign AI efforts. If the model’s combination of regional data weighting, transparent reporting, and efficient long-context design proves useful, it could influence how future open models are built on local cloud infrastructure.

Creati.ai perspective

Soofi S is noteworthy not because Europe produced another open model, but because the team appears to have made a sharper product choice than many public-interest AI projects do. It picked a specific language market, leaned into open documentation, and adopted an architecture optimized for a real serving problem: long-context cost.

The open question is whether those design choices translate into adoption. Enterprise AI buyers care about licensing clarity, benchmark credibility, and operational behavior under messy production workloads. If Soofi S can show reliable performance in German document pipelines and coding assistant scenarios on sovereign infrastructure, it could become more than a symbolic European alternative. If not, it may remain an impressive research release with limited commercial pull.