Soofi S: a German model, and the continental ambition
On June 17, 2026, from Berlin, a consortium of German research institutes announced "Soofi S", presenting it not as just another language model looking for attention, but as the first piece of a family of European AI models. Behind the acronym, Sovereign Open Source Foundation Models, is a project funded by the German Federal Ministry for Economic Affairs and Energy within the IPCEI-CIS initiative, designed to offer businesses, public administration, research centers, and startups a transparent alternative to non-European models. It is a sentence that, read with the right glasses (those of Rowdy Piper in They Live, to be clear, those that reveal what hides behind the advertising sign), does not talk about benchmarks. It talks about power.
To understand why, it is convenient to frame Soofi S immediately in the broader context of the algorithmic war that we have been telling on these pages for months, the one in which Washington and Beijing compete for chips, talents, and access to frontier models, while Europe risks remaining to watch. Soofi S does not change the balance of that game overnight, but it introduces a variable that until now was almost completely missing at the European table—a model of industrial scale built, trained, and distributed entirely on infrastructures of the continent.
Berlin industrializes artificial intelligence
The consortium that built Soofi S says a lot about how Germany is tackling the issue. Not a single laboratory, but a network comprising the Fraunhofer IAIS and Fraunhofer IIS, the German Research Center for Artificial Intelligence (DFKI), the universities of Würzburg, Hannover, and Darmstadt, the Berliner Hochschule für Technik, and the companies Ellamind and Merantix Momentum, all coordinated by the German association for artificial intelligence, the KI Bundesverband. A structure that looks more like an infrastructure consortium—railways or energy—than a Silicon Valley startup.
Jörg Bienert, director of the Association's Center for Sovereign AI, says it bluntly: those who control foundational models control a central part of future digital value creation, and with Soofi they want to build an open base on which businesses and public administration can develop AI applications without permanently depending on non-European models. It is the type of statement that, three years ago, would have sounded like industry propaganda. Today, with Lagarde warning about European delays and with the Commission writing letters of courtesy to Washington asking for access to chips without restrictions, as we told when talking about the algorithmic war, it sounds more like an industrial plan with its papers in order.
The Soofi project goes beyond a single release. Nicolas Flores-Herr, technical manager of the project and team lead at Fraunhofer IAIS, clarifies the objective with an important distinction: Soofi S does not want to be yet another generic chatbot, but a technical base for industrial AI, and what matters is that it works not only in benchmarks, but that it can be deployed reliably, efficiently, and transparently in production. It is a different positioning compared to that of Mistral, which with Devstral 2 points straight to enterprise coding and direct competition with American proprietary models, as we had analyzed a few months ago. Soofi S looks instead at industrial processes, at the analysis of technical and regulatory documents, at agentic systems designed for the German manufacturing fabric, that of the Mittelstands that act as the backbone of the country's economy.
Thirty billion parameters, a hybrid architecture
Let's move on to the numbers, because this is where we measure whether behind the sovereignist rhetoric there is technical substance or just good intentions. The technical report published by the research team describes Soofi S as a Mixture-of-Experts model with a hybrid Mamba-Transformer architecture, of about 31.6 billion total parameters, of which only 3.2 billion are activated for each processed token. Translated for those who do not chew transformer architecture every day, this means that the model behaves, in calculation costs, like a much smaller system, while in overall capacity it approaches a model ten times larger. It is the same principle as a turbocharger on a small engine—power from superior displacement without paying for its consumption in full.
The training, which lasted from late March to mid-May 2026, consumed 27 trillion tokens, distributed over an infrastructure of up to 512 NVIDIA B200 GPUs, the equivalent of 64 DGX B200 nodes. The team deliberately adopted NVIDIA's Nemotron 3 Nano reference architecture without modifications—a pragmatic rather than ideological choice: it means Soofi S can immediately rely on inference stacks already optimized for that family of models, instead of requiring months of integration work by those who want to distribute it.
The practical advantage of this architecture is seen especially on long contexts. While traditional dense models slow down conspicuously when the context window grows, because they have to re-read an increasingly heavy cache at each generated token, Soofi S maintains a generation speed that is substantially constant from 4,000 up to 256,000 context tokens—a detail that is far from academic for those who imagine using it on endless technical documentation or prolonged agentic conversations.
Open, but not yet completely
Here comes the part that deserves more than an enthusiastic headline. The technical report addresses the question of openness with a rigor rarely seen in industry press releases. The authors openly recognize that the definition of open-source artificial intelligence remains contested, with the OSI's Open Source AI Definition 1.0 admitting documented but not redistributable training data, while stricter European proposals require every training token to be redistributable. Soofi S belongs to the first category, not the second.
The weak point, declared with rare honesty in the industry, concerns a specific source: the Genios corpus, granted under commercial license, which represents 1.3% of the actual tokens of the first training phase, and is documented through aggregate statistics rather than redistributed. For the rest, about 99% of the training mixture is publicly obtainable and independently reconstructible. This is transparency that goes beyond the simple publication of weights: the consortium also releases the intermediate checkpoints, the training and evaluation code, and the complete accounting for each data source used in each phase—a level of detail normally found only in reports of efforts like Olmo, not in corporate press releases.
There is, however, an essential clarification, and it concerns precisely the moment we write. The technical sheet published on Hugging Face specifies that the checkpoint currently available is a preview in the closed beta phase with selected partners, not yet an open release, and that the final model will be released openly under a permissive license, without bound access. In other words, the promise of openness is written in black and white, but it is not yet verifiable reality by anyone. For a project that makes transparency its identity argument, this gap between declared intention and actual availability deserves to be kept under observation, and it is the reason why this article focuses on the political meaning of the project rather than on a definitive technical judgment: we wait for the release of the weights for a field test, and at that point we will return to the topic with hand on keyboard.
German, English, and the Italian that waits
Those who read this piece hoping for a European model capable of speaking fluent Italian will be partly disappointed. The technical sheet describes Soofi S as a bilingual German-English model, not generically multilingual, with a design choice declared since the abstract of the report. German receives a privileged treatment that goes far beyond the average of international models: in the widespread pre-training phase, the German language represents 7.2% of the mixture, against the typical 5% of the reference recipe, while in the final refinement phase it rises to 15.3%, more than triple the standard.
For languages like French, Spanish, and Italian, the technical sheet is honest to the point of caution: they are covered in a limited way, mainly through small subsets of multilingual fine-tuning. It is not a design flaw, it is an explicit choice, and probably also the most sensible one from an industrial point of view: building a truly competitive model in two languages, instead of a mediocre one in twenty. But it means that, for the Italian public, enthusiasm must be tempered: Soofi S is a highly relevant political and economic case study for Europe as a whole, not (yet) a tool designed for the Italian linguistic market. Any evaluation of its capabilities in Italian must be verified with dedicated tests when the weights are accessible, not taken for granted on the basis of the results in German.
Where it convinces, and where it doesn't
The technical report compares Soofi S with fifteen open models in two distinct groups: that of genuinely open-source models (Alia, EuroLLM, Apertus, Olmo 3) and that of larger open-weight models (Qwen, Ministral, Gemma), besides the direct comparison with Nemotron 3 Nano, the reference architecture. In the first group, that of peers, the results are clear: Soofi S obtains the highest aggregate score in English, with a margin of 2.8 points over Olmo 3, and the highest aggregate score in German, with a margin of 6.3 points over Apertus, resulting in the strongest fully open model of the comparison in both English and German.
On coding, historically an Achilles' heel for European projects, the gap is even more marked: Soofi S leads four of the five programming benchmarks considered, surpassing the best open-source opponent by 10.8 points on HumanEval and 13.4 points on MBPP in German. Numbers that, if confirmed on a public release verifiable by anyone, would significantly scale down the narrative that Europe only produces regulations and never code.
The picture changes, and it is right to say it with the same clarity, when the comparison extends to larger open-weight models. Against Qwen3.5, Ministral 3 14B, and Gemma 3 27B, Soofi S is not the strongest model on aggregate scores, but remains competitive with larger dense bases, with 70.1 points on the English aggregate against 70.3 for Gemma 3 27B and Ministral 3 14B—an almost imperceptible gap considering that Soofi S activates a fraction of its rivals' parameters. Compared to the direct architectural progenitor, Nemotron 3 Nano, the leap is clear everywhere, a sign that the German-English data recipe built by the consortium produces a measurable and not only theoretical benefit.
Munich, Deutsche Telekom, and the race for compute
If there is one element that really moves the needle of the geopolitical compass, it is the infrastructural choice. Soofi S was built entirely on the Industrial AI Cloud operated by Deutsche Telekom in Munich, which went into operation in February 2026, and not on computing capacity rented from American hyperscalers. The technical report states it unambiguously: training on German soil, under European operational and data protection requirements, is an integral part of the model's sovereignty, and did not take place on non-European hyperscale infrastructures.
It is a detail that is worth more than many political announcements, because it shifts the focus from the model to the chain. It is one thing to publish weights downloadable by anyone, it is another to control the entire production cycle, from the chip to the network, passing through cooling and energy. On this front, the project also claims an environmental profile consistent with the rest of the European narrative: the Munich facility is powered entirely by renewable energy, cooled with water from the Eisbach canal, and integrated into a waste heat recovery system for the surrounding district. An almost bucolic detail, compared to Texan data centers that drain aquifers, but one that tells a different infrastructural philosophy.
However, there remains the structural limit we had already highlighted when talking about Mistral: even the Industrial AI Cloud, as much as it is operated on German soil, runs on NVIDIA GPUs, Californian by design and Taiwanese by manufacture. European computational sovereignty, in this historical moment, means controlling where the silicon runs, not yet who produces it. It is a distinction worth keeping in mind every time we read the word "sovereign" associated with a European AI project, whether French or German.
Why Europe insists on its own models
Soofi S is not born in a vacuum. It arrives after months in which technological sovereignty has passed from a conference concept to a declared priority in Christine Lagarde's speeches, and after American restrictions on AI chips split Europe into two groups of countries with different access to the most advanced GPUs, as we reconstructed by analyzing the ongoing algorithmic war between Washington and Beijing. In that context, an industrially scaled model trained entirely on European infrastructures is not just a technological product, it is a negotiating argument.
Luciano Floridi's vision, which we recounted when talking about Mistral, imagines a fourth global AI pole, distinct from the United States and China, built on the meeting between open source and smart regulation. Soofi S fits into that trajectory with a different accent compared to Mistral: where the French company focuses on a hybrid commercial model, differentiated licenses, competitive pricing, international enterprise clients, the German consortium focuses on a semi-public good, funded with state and European funds, designed explicitly for German industry and public administration even before the global market. They are two complementary rather than competing strategies, and their co-existence says something interesting about the way Europe is, painfully, learning to play on multiple tables simultaneously.
Then there is the less visible but probably more decisive dimension in the long term: that of data. The training corpus includes German journalistic sources granted under commercial license, European academic corpora, technical and regulatory texts. It is a different information ecosystem from the one, dominated by GitHub, Reddit, and Stack Overflow, on which American models feed. If Europe really wants to build a culturally distinct alternative, and not just a technically equivalent one, the composition of the data matters as much as the model's architecture, perhaps even more. In this sense, Soofi S recalls, in approach if not in declared ambition, certain independent Japanese projects that in recent years have insisted on the same idea: a model is never neutral with respect to the culture that trained it, and those who control that culture also control, in part, the imagery that the model returns.
Access still closed, a second chapter to be written
It is worth being clear on one point, before closing. Everything we have told so far is based on a rigorous technical report, on an institutional press release, and on a model card published by those who built the model. They are authoritative sources, but they are still the voice of those who have an interest in telling their work well. Soofi S, in its current state, is distributed only to a selected number of partners in closed beta, it is not downloadable or queryable by whoever writes nor by the readers of this article.
For this reason, we chose to treat this news especially in its political and strategic dimension, rather than as a technical review, with the critical cut that seemed to us more honest given the still partial nature of the available information. The day the weights are actually public, as promised by the consortium, we will return to Soofi S with a real field test: reasoning on long texts, tests in Italian, direct comparison with Devstral 2 and with other European open models, verification of declared performances on accessible hardware. Until then, every number cited in this piece must be read as a statement of those who built the model, not as an independent verification.
Turning point or well-constructed political signal?
We remain with the question that, ultimately, matters more than any benchmark. Is Soofi S really the first brick of an autonomous European artificial intelligence chain, or is it above all a political signal constructed with notable technical care? The honest answer is that, in this phase, both are true together, and it is by no means a contradiction.
It is a political signal because it arrives at the exact moment when Europe needs to prove to itself, before even to the markets, that it can do something more than regulation. It is at the same time a real technical brick, because behind the announcement there is a training run of 27 trillion tokens, a modern architecture chosen with solid engineering criteria, a document transparency commitment that few projects, European or not, have the courage to maintain to the end, including admitting their own limits.
It is also worth asking what would change if this effort were not German, but truly European. A consortium that put together French computing power, Italian and Spanish linguistic data and skills, Brussels' regulatory experience, and German industrial capital would have a scale difficult to achieve by a single state, even the most determined. The risk, however, is well known to anyone who has followed major community projects, from Galileo satellites to sixth-generation fighters: more countries means more vetoes, more mediations on the training language, more time lost dividing industrial fallouts instead of shipping models. Soofi S proves that a single country, with clear direction and targeted funds, can move fast. A consortium of twenty-seven could do something bigger, assuming Europe finally learns to decide together without slowing each other down.
The real test will come when the closed beta ends, and anyone can download the weights, query the model, and truly measure its capabilities outside the consortium's labs. Only then will we know if Soofi S is the first chapter of a European story that lasts over time, or if it will remain a case studied in digital sovereignty conferences without ever becoming the silent infrastructure on which the continent's industrial applications truly run. Until that day, the finger points to Germany, but the moon, as always in this game, remains the whole of Europe.