A working review of Soofi S, the open ~30B German consortium LLM. What it nails on efficiency and German, where it stops as a base model, and who it fits.
Soofi S is a standout open release: an efficient ~30B mixture-of-experts model that activates only ~3.2B parameters per token, tops fully open models on both German and English, and was built in Germany on sovereign infrastructure. Judged as what it is — an open, German-strong base model — it is excellent and genuinely useful. Two honest caveats hold it back for most creators: it is text-only, and the first release is a base model aimed at research and fine-tuning, not a ready assistant, with the license still being finalized. Score it high for efficiency and German; look elsewhere if you came to produce and ship content.
Most coverage of Soofi S is either a "Germany finally has its own model" flag-wave or a benchmark table without context. This review is neither. We build a content engine and read model cards for a living, so the goal is to tell you what Soofi S is genuinely good at, where it stops, and — because people arrive at this question sideways — whether it can power a content operation.
Short version up top: Soofi S is an impressive open model. Released on July 13, 2026 by the Soofi project — a German consortium coordinated by the KI Bundesverband with Fraunhofer institutes, the DFKI, several universities, and the companies Ellamind and Merantix Momentum, funded by the German Federal Ministry for Economic Affairs and Energy under the IPCEI-CIS program — it pairs an unusually efficient architecture with a deliberate German focus. About 31.6B total parameters but only ~3.2B active per token, a hybrid Mamba-2/attention design, trained on roughly 27 trillion tokens on Deutsche Telekom's sovereign cloud in Munich, and it reports the top aggregate open-model scores on German and English, ahead of OLMo 3 32B and Apertus 70B.
The honest catch is scope and readiness. Soofi S is a text-and-reasoning model — it renders no images, no video, no audio, and it publishes nothing. And the initial release is a base model, positioned for further post-training and research rather than plug-and-play assistant use, with the release license not fully finalized at launch. None of that is a flaw — it set out to be an efficient open base layer, not a finished application — but it is the thing to understand before deciding it fits your workflow.
This review covers what Soofi S actually is in 2026, how its efficiency and German strength stack up, where it is honestly the wrong tool, and who should use it versus who should keep looking.
Soofi S is an open large language model from the Soofi project, a government-funded German research consortium. Its full name is Soofi-S-30B-A3B: about 31.6 billion total parameters, but a mixture-of-experts design that activates only roughly 3.2 billion per generated token, so its compute cost sits closer to a 3B model than a conventional 30B one. Architecturally it is a hybrid — Mamba-2 state-space layers mixed with a small number of standard attention layers (of its 52 layers, only about 6 maintain a KV cache) — which keeps throughput high on long inputs; the consortium reports on the order of eight times more tokens per second per GPU than dense 14–24B models at a 40,000-token context. It was trained on about 27 trillion tokens across three phases, with a long context (up to a million tokens; effective through roughly 256k). What sets it apart is language balance. Where most open models treat non-English as an afterthought, Soofi S deliberately raised the German share from 7.2% of an early phase to 15.3% later, and reports the top aggregate score among fully open models on both German and English, with strong German coding, science, and general-knowledge results. English and German are primary; French, Italian, and Spanish have lighter support. The weights are published on Hugging Face, but the release license had not been finalized at launch (a small share of training data — about 1.3%, from commercially licensed newspaper archives — keeps it from meeting the strictest European open-data definitions), and the initial checkpoint is a base model intended as a foundation for instruction tuning, preference tuning, and domain adaptation rather than direct end-user deployment.
The clearest fit is a German enterprise, public body, or research team whose requirement is sovereignty, efficiency, and strong German: anyone who needs to run and audit a capable model on its own infrastructure, or a builder who wants an efficient open base model to fine-tune without vendor lock-in. Its MoE efficiency makes it attractive where GPU budget matters, and its German strength is a genuine differentiator over models that treat the language as an afterthought. It is the wrong tool for someone whose actual output is published content — video, images, carousels, social posts — because producing and distributing that is entirely outside what a text model does, and for non-technical users who want a hosted, log-in-and-go assistant rather than a base model to tune and operate.
| Dimension | Score | Why |
|---|---|---|
| Efficiency (MoE / throughput) | 4.7 / 5 | Activating ~3.2B of ~31.6B and reporting ~8× the tokens/sec/GPU of dense 14–24B models on long context is the standout — it runs far cheaper than its size suggests. |
| German + English quality | 4.5 / 5 | Reports the top aggregate open-model scores on both languages, ahead of OLMo 3 32B and Apertus 70B, with strong German coding and science results. |
| Openness & availability | 4.0 / 5 | Weights are published, but the release license was not finalized at launch and ~1.3% of data is commercially licensed, so it is not "fully open" in the strictest sense. |
| Long-context handling | 4.2 / 5 | Trained up to a million tokens and effective through roughly 256k, with throughput that holds up on long inputs thanks to the hybrid design. |
| Sovereignty & data posture | 4.4 / 5 | Built and trained in Germany on Deutsche Telekom's sovereign, renewable-powered cloud, government-funded and institution-backed — a strong regulated/public-sector story. |
| Model quality & reasoning (general) | 4.0 / 5 | Strong for its class and efficiency, but not positioned as a frontier-leading reasoner against the top closed models. |
| Readiness / ease of use | 3.0 / 5 | The first release is a base model aimed at post-training and research; most production use needs fine-tuning or a tuned variant, not download-and-go. |
| Ecosystem & tooling | 3.2 / 5 | Brand-new and young — a Hugging Face model card and consortium materials, without the mature hosted tooling of established open families. |
| Content / social media production | 1.0 / 5 | Not the product. No image, video, audio, captions, or design output. Out of scope by design. |
| Multi-platform publishing | 1.0 / 5 | Soofi S produces text; it does not post. There is no scheduler and no platform integration. |
Soofi S has no license price. The weights are free to download, so the cost question is really "what does it cost to run and make useful." Self-hosting means GPU and inference infrastructure — and, because this is a base model, likely a fine-tuning step to get clean instruction-following before it is production-ready. The upside is that its mixture-of-experts efficiency, activating only ~3.2B parameters per token, meaningfully lowers the inference bill relative to a dense 30B model, which is exactly the point of the design.
For the sovereignty and research use cases Soofi S targets, that economics is right: an efficient, self-hostable, German-strong model is enormous value when control, cost, and German quality are the requirements. The caveat is the license — it was not finalized at launch, and about 1.3% of the training data comes from commercially licensed newspaper archives, so anyone planning commercial use should confirm the current terms on the model card rather than assume a permissive license.
The honest framing on value is that Soofi S is priced like what it is: open infrastructure, and unusually cheap to run for its capability. It is not priced or built as a content-marketing tool, and no amount of inference budget adds rendering or publishing. If your spend is meant to produce and distribute content, you are comparing the wrong line item.
| Use case | Fit | Why |
|---|---|---|
| Sovereign / on-prem deployment for German or EU regulated use | Strong | Self-hostable and built in Germany on sovereign infrastructure, government-funded and institution-backed — squarely its purpose. |
| German-first text generation and translation | Strong | Deliberately German-weighted and tops fully open models on German benchmarks — a genuine differentiator. |
| Fine-tuning an efficient open base model | Strong | MoE efficiency and long context make it a strong, low-cost foundation to post-train for your own product without lock-in. |
| Drafting copy, scripts, and summaries | OK | Capable text generation, but as a base model output needs tuning or a human pass, and it is one input to a content workflow, not the whole thing. |
| Ready-to-use hosted chatbot out of the box | Weak | The initial release is a base model, not an instruction-tuned assistant; expect to fine-tune or wait for a tuned variant. |
| Producing short-form or avatar video for social | Weak | No video generation of any kind. Entirely outside Soofi S's scope. |
| Brand-consistent content across formats | Weak | No persona or brand-voice system, no design layer. It generates text but does not govern a content voice or render media. |
| Scheduling and publishing across platforms | Weak | No publishing layer and no scheduler. Soofi S produces text, not posts. |
If you arrived at this review wondering whether Soofi S can run your content operation, the honest answer is no — and that is a category question, not a criticism. Soofi S is a foundation model: efficient, German-strong, sovereign open infrastructure. It has no renderer, no caption engine, no design layer, and no scheduler, because it was never meant to be a content tool — and as a base model, it is not even a finished assistant yet. Scoring it as a content engine would be unfair to a model that is genuinely excellent at its actual job.
Kompozy sits at the layer above. Where Soofi S stops at text, Kompozy turns an idea or a draft into 18 content formats — persona and avatar video, carousels, quote cards, infographics, blogs, newsletters, and platform-native posts — holds one brand voice through a Persona Brief, and schedules and publishes across nine platforms plus email and blog. It runs that generation on managed Claude and OpenAI models, so there are no GPUs to operate and nothing to fine-tune. The two are not rivals: a German sovereignty-minded team could draft in its own Soofi S deployment, keeping the German-first text on its own infrastructure, and let Kompozy produce and ship the finished content. Use Soofi S for the efficient open model layer it is built for, and a content engine for the content.
Soofi S is an open large language model released on July 13, 2026 by the Soofi project, a German research consortium coordinated by the KI Bundesverband. It has about 31.6B total parameters but activates only ~3.2B per token via a mixture-of-experts design, uses a hybrid Mamba-2/attention architecture, and tops fully open models on aggregate German and English benchmarks.
For efficiency, German strength, and sovereignty — yes, it is one of the strongest open releases available and cheap to run for its size. It is not worth adopting for content production, because it generates no media and publishes nothing, and its first release is a base model that needs fine-tuning before it is a smooth assistant.
Very good for an open model. It was deliberately trained with a much higher German share than typical open models — rising from 7.2% to 15.3% of the mix across phases — and its pretraining report claims the top aggregate German and English scores among fully open models, ahead of OLMo 3 32B and Apertus 70B.
The weights are free to download, so your real cost is the GPU/inference infrastructure to self-host it, plus any fine-tuning to make the base model production-ready. Its MoE efficiency (only ~3.2B active parameters per token) lowers that inference bill relative to a dense 30B model, but confirm the release license before commercial use.
No. Soofi S is a text-and-reasoning model. It does not render video, generate images, write captions, or post to any platform. To turn its text into published content you need a content engine like Kompozy.
Not out of the box. The initial release is a base model aimed at research and further post-training (instruction tuning, domain adaptation). It is efficient to run, but expect to fine-tune it or use a tuned variant for smooth, production-grade assistant behavior.
Because it was built and trained in Germany — on Deutsche Telekom's Industrial AI Cloud in Munich, funded by the German government and coordinated by a German consortium — and its open weights let institutions run and audit it on their own infrastructure rather than depending on a closed foreign provider. That independence is the sovereignty argument.
Kompozy, without question. Soofi S produces text and nothing else; Kompozy generates video, images, carousels, blogs, and newsletters and publishes them across platforms. Use Soofi S as an efficient open model layer — for German-sovereign teams, even as a drafting source — and Kompozy to produce and ship the content.