Soofi S 30B-A3B, published on arXiv (DOI) on July 10, 2026, is a 30-billion-parameter Mixture-of-Experts model that activates only 3B parameters per token, trained end-to-end on German infrastructure for German and English. It is the strongest case yet that European “sovereign” AI can mean auditable training provenance rather than a model card that merely promises data residency. Sovereignty buys control, not capability, and keeping the weights current is the part nobody has budgeted for.
What does “sovereign” actually guarantee?
A sovereign model, as Soofi S defines the term, ships open weights alongside a full accounting of where every training token came from, which infrastructure trained it, and how it was evaluated, rather than a residence claim about where inference runs.
The release includes the weights, selected intermediate checkpoints, per-source data accounting, hyperparameters, and the training and evaluation code, all under permissive open-access terms (arXiv:2607.09424). The entire pretraining run executed on the German Industrial AI Cloud operated by Deutsche Telekom in Munich. That distinction is what compliance teams should care about: data residency says your prompts stay in Germany; auditable provenance says the model’s knowledge base is inspectable down to the source.
The label is doing real work here, because “sovereign AI” has become a phrase that means whatever the speaker needs it to mean. A vendor can claim sovereignty by routing inference through a Frankfurt endpoint while the underlying weights and training data remain opaque. Soofi S’s claim is harder to fake: open weights let a third party reproduce behavior, and per-source data accounting lets an auditor trace what the model actually learned from. The bar is not where the compute sits but whether the training pipeline is open to inspection.
How is the architecture built, and why does 3B active matter?
Soofi S is a hybrid Mamba-Transformer Mixture-of-Experts model with 30B total parameters that activates roughly 3B per token, which means inference cost tracks the active parameter count rather than the full weight footprint.
The Mixture-of-Experts routing sends each token through a small subset of expert subnetworks, so a 30B model can run with the latency and memory profile of something closer to a 3B dense model. The hybrid Mamba component brings state-space-model dynamics into the sequence path, where inference time grows roughly linearly with sequence length rather than quadratically as in pure attention. For a German-hosted deployment running long-context workloads, that is the difference between a model that fits a fixed memory budget as context grows and one that does not. The architecture is tuned for the cost structure of running compliant inference on European GPUs rather than for topping a leaderboard.
The efficiency story is the part that actually transfers to operations. Matching a dense model’s quality at a fraction of the active parameter count is what makes self-hosting defensible on a cost basis; without it, the sovereignty argument collapses into “we run a slower model in Munich.”
How was the training data weighted, and where did it run?
Soofi S was pretrained on roughly 27 trillion tokens with deliberately up-weighted German content, with the full run executed on Deutsche Telekom’s German Industrial AI Cloud in Munich (arXiv:2607.09424).
The German up-weighting is the sovereignty thesis expressed in training data: a model that sees proportionally more German than a global frontier corpus would feed it, tuned for a linguistic region rather than an English-default average. For German-language legal, regulatory, and industrial text, that weighting is the difference between a model that handles the domain natively and one that translates through an English pivot and loses register.
How does it benchmark against other open models?
Among 17 open base models tested, Soofi S achieves the best code aggregates in both German and English and matches dense 14-27B models on aggregate benchmarks, while outperforming Olmo 3 32B and Apertus 70B among fully open models on English and German evaluation scores (arXiv:2607.09424).
That positions Soofi S at the top of the fully-open European field. It does not position it against frontier APIs. Matching a dense 14-27B model with 3B active parameters is a genuine efficiency result; matching GPT-class or Claude-class capability it is not, and the paper does not claim it is. The preprint does not name specific prior European models such as LeoLM or OpenGPT-X, so the head-to-head lineage against earlier European efforts is not fully transparent from the abstract alone. Readers comparing Soofi S to a specific predecessor should verify the pairing against the paper’s evaluation tables rather than the abstract.
Soofi S wins its lane, and its lane is open European models. The frontier is a different lane, and sovereignty does not relocate the model into it.
Why are open weights cheap to adopt but expensive to keep current?
Open weights make a sovereign stack nearly free to adopt and costly to sustain, because the institution that adopts them inherits the burden of retraining whenever the frontier moves.
This is the second-order effect the sovereignty discourse tends to skip. A European enterprise can download Soofi S, host it on Telekom’s cloud, pass a data-residency audit, and ship a product. What it cannot do is stop there. Frontier labs retrain on fresh web crawls, new code, and new instruction data on a rolling basis; an open sovereign model is a snapshot of one training run on one corpus. The 27-trillion-token pretraining that produced Soofi S is the floor of what reproducing it costs, and reproducing it is the minimum to stay within a year of where the frontier was when the snapshot was taken.
That retraining tax is the real constraint on European AI independence, and it does not fall on the model’s authors. It falls on whichever institution decides to keep the weights competitive. Deutsche Telekom can ship a sovereign model once; keeping a sovereign model current is an ongoing compute commitment that most adopters are not structured to make. Open weights lower the adoption cost to roughly zero and raise the maintenance cost to roughly the original training cost, paid repeatedly.
What do practitioners gain, and what do they give up?
Practitioners gain a compliant, auditable, German-hosted stack with long-context inference efficiency; they give up frontier capability and accept ongoing maintenance as their own problem.
The gains are concrete and operationally useful. The compliance posture is real: open weights plus per-source data accounting plus German-hosted compute is an audit you can actually walk a regulator through. The long-context efficiency from the Mamba hybrid is real: bounded cache growth makes long-document German regulatory and legal workloads deployable without the memory tax of a dense attention model. The licensing is real: permissive open-access terms mean an institution can modify, fine-tune, and redistribute without renegotiating a vendor contract.
The losses are equally concrete. The capability gap to frontier APIs is not closed by sovereignty; it is bracketed by it. The closed-data gap remains: aggregate data accounting is not the same as a fully disclosed corpus, and the commercially licensed sources the paper does not enumerate are exactly the sources a compliance audit would want named. And the maintenance burden is transferred wholesale to the adopter.
The trade is control for currency. For an institution whose constraint is compliance rather than capability, that trade is rational. For an institution whose product depends on frontier reasoning, it is not.
Frequently Asked Questions
Does Soofi S support languages beyond German and English?
The paper specifies German and English as the target languages with German-upweighted pretraining, but does not document systematic evaluation on other European languages. Practitioners deploying for French, Italian, or Spanish workloads should treat cross-lingual performance as untested rather than implicitly supported.
What specific compliance gaps remain despite open weights?
The paper reports data accounting in aggregate and does not enumerate which commercially licensed corpora remain closed. For a license-compliance audit, that gap matters because an auditor needs to trace specific sources, not just aggregate statistics. The model ships with inspectable provenance but not fully disclosed sourcing.
How does the 27-trillion-token training scale compare to frontier runs?
The 27 trillion token figure represents a large open-model training run, but frontier labs have largely stopped publishing comparable token counts. Without a direct baseline, 27T is a statement about this specific run rather than evidence of parity with frontier pretraining budgets, which may run substantially higher.
What infrastructure is required to run Soofi S efficiently?
The hybrid Mamba-Transformer MoE architecture activates roughly 3 billion parameters per token, allowing inference with the memory profile closer to a dense 3 billion parameter model despite 30 billion total parameters. This requires European GPU infrastructure optimized for the routing and state-space-model components, not just general-purpose compute.
How frequently would retraining be required to stay competitive?
The model is a snapshot of one training run on one corpus from July 2026. Frontier labs retrain on fresh web crawls, new code, and new instruction data continuously. Adopters must budget for repeated pretraining runs at similar scale to stay within a year of frontier capability, not a one-time download cost.
