Soofi S makes Europe’s open-model race about receipts, not just weights

AI reporting — Deena Turner

The plain-English takeaway: a German-backed research consortium has published Soofi S 30B-A3B, a German-English open foundation-model project that matters less because it claims to “win” a leaderboard and more because it tries to make the model’s training history auditable. The team says the model activates only about 3.2 billion of its roughly 31.6 billion parameters per token, was trained on about 26.68 trillion tokens, and will be released with weights, selected checkpoints, training code, evaluation code, and unusually detailed data accounting. That is consequential for readers because the practical question in open AI is no longer simply “Can I download the weights?” It is “Can I understand what was trained, how it was trained, what it costs to serve, and where the claims stop?”

This is not a story about a model suddenly overtaking the closed frontier. The Soofi team does not demonstrate that. It is also not a finished public deployment. The project page says model repositories are currently gated during a beta phase and that the models will be freely available after that phase ends. The evidence today supports a narrower but important conclusion: Soofi S is a serious European open-model release package, with primary documentation that raises the transparency bar while still leaving independent replication and real-world deployment performance unproven.

What was released

The Soofi project page describes Soofi S 30B-A3B as a “sovereign, open-source Mixture-of-Experts hybrid Mamba–Transformer foundation model for German and English.” The report lists the consortium as including KI Bundesverband, DFKI, Fraunhofer IAIS, Fraunhofer IIS, Technische Universität Darmstadt, Universität Würzburg, Berliner Hochschule für Technik, L3S Research Center, Lamarr, ellamind, hessian.AI, and Merantix Momentum. It says the consortium was coordinated by the KI Bundesverband and funded by Germany’s Federal Ministry for Economic Affairs and Energy in the context of IPCEI-CIS and 8ra.

The model is described as a 31.6-billion-parameter system with about 3.2 billion active parameters per forward pass, or about 3.6 billion including embeddings. The architecture, according to the technical report, uses 52 layers: 23 Mamba-2 sequence-mixing layers, 23 granular mixture-of-experts layers, and 6 grouped-query attention layers. The point of that design is not just parameter-count theater. The project argues that only the six attention layers keep a key-value cache, so long-context serving should avoid some of the memory growth that slows full-attention Transformer models.

The training claim is specific: the report says Soofi S was pretrained on 26.68 trillion tokens, deliberately up-weighting German alongside English. The public project page rounds that to roughly 27 trillion tokens and says German reached up to 15.3% of the mixture. The report also says training ran from March 24 to May 13, 2026, on up to 512 NVIDIA B200 GPUs in Deutsche Telekom’s German Industrial AI Cloud in Munich.

That makes the release part model story, part infrastructure story, and part governance story. The consortium is not only saying it built a German-English model. It is saying the model was trained end-to-end on a European cloud environment under European operational and data-protection requirements. That is a claim about where capability is produced, not only how it benchmarks.

What is available now — and what is not

The project page lists several artifacts: base weights and selected intermediate checkpoints, training code and reproducible data-construction scripts, evaluation code, a Weights & Biases dashboard for the full run, exact per-source token accounting for all three phases, and documentation for commercially licensed sources in aggregate form.

But availability needs careful wording. The same page says that, during the current beta phase, the model repositories are gated and users must accept Hugging Face access conditions before downloading. It also says the models will be freely available without access request once the beta ends. That means this is not the same as an instantly unrestricted release today. It is a documented release package with gated model access at the moment.

The training-code repository is public and describes scripts, documentation, and debugging material for Soofi pretraining, including preprocessing code for corpus construction, filtering, tokenization, and mixture assembly, plus pretraining and midtraining configuration directories. The repository README says the current code release contains the complete codebase used for training Soofi S, a 30B model based on NVIDIA’s Nemotron 3 Nano architecture. That is meaningful evidence for the project’s transparency claim, though a full reproducibility audit would still require running or inspecting the code, verifying data availability, and checking whether every advertised artifact is present and usable under its license.

The release also comes with a license caveat. The project page lists the Hugging Face Space license as Apache-2.0, while the model page visible on Hugging Face labels Soofi-S-Base with “other” as its license. The technical report says Soofi S “will be released under highly permissive, open-access terms,” but today’s safest reader-facing wording is that the project says artifacts are being released under permissive terms, while exact license status should be checked on each artifact before commercial or regulated use.

The benchmark claims, with the brakes on

The Soofi project page says the model reaches English aggregate 70.1 and German aggregate 79.1, ahead of fully open models it compares against: Olmo 3 32B at 67.3 English and 69.2 German, and Apertus 70B at 62.4 English and 72.8 German. It also reports code results of HumanEval 73.8, MBPP 70.2, and MBPP-DE 84.2; mathematics results of GSM8K 86.1, GSM8K-Platinum-DE 87.1, and Minerva-500 79.4; and German benchmark results including GLP-DE 88.8, INCLUDE-DE 61.2, and ARC-Challenge-DE 92.3.

Those numbers are publishable as the team’s reported evaluation results, not as independent fact about superiority. The important sentence in the report is methodological: it says Soofi S was evaluated against 15 open models of comparable or larger active size across parallel English and German benchmarks, and it defines a capability index that averages five benchmark groups after normalization. That gives readers a better map of what is being measured, but it does not remove the usual benchmark-theater risks.

First, aggregate scores hide task composition. A German-English aggregate can reward a model for doing well on the selected mix without proving broad performance across every German-language professional use case. Second, code benchmarks such as HumanEval and MBPP are useful signals, but they are not the same as maintaining a messy production repository, following a security policy, or handling ambiguous product requirements. Third, comparisons to other models depend on evaluation harness versions, prompting, sampling settings, contamination controls, and whether the comparison models were run locally or copied from official reports. The project deserves credit for releasing evaluation materials, but independent reruns are the next check, not a decorative footnote.

The serving-efficiency claim is more practically interesting than the leaderboard claim. The project page says Soofi S reaches 4.82 thousand aggregate decode tokens per second per GPU at 40,000-token context, batch size 32, on a single B200 using vLLM. It describes that as 9.2 times Ministral 3 14B, with near-flat decode throughput from 4,000 to 256,000 tokens. The report explains the mechanism: because only six of 52 layers keep a KV cache, the per-sequence cache stays closer to constant as context grows.

That is plausible as an architectural advantage, but still a reported result. It is not a cloud bill, a production latency SLO, or a user-facing deployment audit. Throughput can move with kernel versions, quantization, batching strategy, prompt length distribution, and hardware utilization. For builders, the right conclusion is not “Soofi is faster everywhere.” It is “this architecture makes a specific, testable serving-cost claim that deserves independent measurement.”

Why this matters beyond one model

Europe’s “sovereign AI” debate often gets trapped between two weak stories. One says local models are mostly symbolic because the frontier is elsewhere. The other says sovereignty itself is a capability substitute. Soofi S is useful because it makes both shortcuts harder.

The capability evidence is still bounded. The model is a base model, not a fully demonstrated consumer assistant or enterprise agent. The public materials do not prove that it beats closed frontier systems. They do not show deployment at national scale. They do not answer every safety, privacy, or abuse question. But the release does show a concrete path for European institutions that want more than procurement dependence: choose an efficient architecture, train on sovereign infrastructure, document the data mixture, publish the recipe, and let outsiders inspect the parts that can legally be shared.

For companies, the interesting part is the deployment math. A model that activates roughly 3.2 billion parameters out of 31.6 billion and keeps the attention cache limited could be attractive for long-context workloads where dense models become expensive. That does not mean a procurement team should swap models after reading a benchmark table. It means the evaluation plan should include long-context throughput, German task quality, license review, data-governance review, and failure testing on the organization’s own workloads.

For researchers, the release is a reminder that openness is a stack, not a badge. Weights are one layer. Data accounting is another. Training code, intermediate checkpoints, hyperparameters, evaluation harnesses, and logs are additional layers. Soofi’s strongest claim is not that every number is independently settled today. It is that the team is exposing enough of the system for outsiders to check the numbers.

For policy readers, the release sharpens the question behind public AI funding. If governments fund foundation-model projects, the public-interest return should not be only a press release or a national-branded chatbot. It should include durable public artifacts: reproducible recipes, transparent data documentation, permissive research access, and measurements that can be challenged. Soofi S appears designed around that standard. Whether it meets it fully depends on the gated beta ending, the promised artifacts staying available, and independent users being able to reproduce the major claims.

Claim ledger

| Claim | Evidence | Confidence | Caveat |
|---|---|---:|---|
| Soofi S is a German-English open foundation-model project coordinated by a German consortium. | Soofi project page and technical report list the institutions, coordinator, funder, and model scope. | High | “Open” still depends on artifact-level access and licenses. |
| The model has about 31.6B total parameters and about 3.2B active parameters per pass. | Technical report architecture section and project page at-a-glance summary. | High for reported specification | Not independently inspected from weights in this review. |
| Training used about 26.68T tokens and ran on up to 512 NVIDIA B200 GPUs in Deutsche Telekom’s Munich cloud. | Technical report and project page. | Medium-high | This is a project-reported training account, not an external audit. |
| The model currently has gated beta access. | Hugging Face project page note on gated repositories. | High | Access status can change; readers should check the model page before acting. |
| The reported benchmark aggregates put Soofi S ahead of selected fully open models in English and German. | Project page and technical report evaluation section. | Medium | Independent reruns are needed; aggregate scores depend on benchmark mix and harness details. |
| The long-context throughput claim is technically plausible and important. | Project page reports 4.82k aggregate decode TPS/GPU at 40K context and explains cache behavior; report describes the hybrid Mamba–MoE architecture. | Medium | Production latency and cost require independent measurement under real workloads. |

What to watch next

The next useful checks are straightforward. First, does the beta end with ungated access to the promised base weights and selected checkpoints? Second, do outside evaluators reproduce the English, German, code, math, and serving-efficiency results using the released harnesses? Third, are the data-accounting files complete enough for researchers to understand excluded sources, licensed sources, and language balancing? Fourth, do real users find the model strong on German professional tasks that benchmarks only approximate: legal drafting, public-sector forms, technical support, medical-adjacent caution, and enterprise search?

Until those checks are done, Soofi S should be treated as a substantial, well-documented release — not a settled leaderboard crown. The reason to pay attention is not that one more model says it is best. It is that this one gives readers more receipts than usual, and in AI, receipts are the beginning of trust.

Sources

• Soofi S project page on Hugging Face
• Soofi S Pretraining Report v1.0 PDF
• Soofi Project organization on Hugging Face
• Soofi-S-Base model page on Hugging Face
• Soofi Pretraining code repository