An open ~30B German-and-English language model from a German research consortium, built for sovereign AI and efficient enough to run near a 3B compute cost.
Last verified · 2026-07-15 · by Moe Ameen
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 (the German AI Association). Its members include the Fraunhofer institutes IAIS and IIS, the DFKI, TU Darmstadt, the University of Würzburg, the L3S Research Center, Berlin University of Applied Sciences, and the companies Ellamind and Merantix Momentum. The project is backed by the German Federal Ministry for Economic Affairs and Energy under the European IPCEI-CIS program. The framing is "sovereign AI": a capable model, built and trained in Germany, that public institutions and companies can download, inspect, and run themselves rather than renting from a closed US or Chinese provider.
The model's headline is efficiency. 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 parameters per generated token, so its compute cost sits closer to a 3B model than a conventional 30B one. The architecture 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 is what lets throughput stay high on long inputs. According to the consortium, at a 40,000-token context with parallel requests it can generate on the order of eight times more tokens per second per GPU than dense models in the 14–24B range. It was trained on about 27 trillion tokens across three phases, on Deutsche Telekom's Industrial AI Cloud in Munich (powered by renewable energy). It was trained with a long context (up to one million tokens) and tested as effective through roughly 256,000 tokens.
The differentiator is language balance. Where most open models treat non-English as an afterthought (in one common reference recipe, all non-English languages together are about 5% of the mix), Soofi S deliberately raised the German share from 7.2% in an early phase to 15.3% later. The result, per its own pretraining report, is the top aggregate score among fully open models on both German and English benchmarks — ahead of prior leaders like OLMo 3 32B and Apertus 70B — with strong German coding, science, and general-knowledge results. English and German are primary; French, Italian, and Spanish have more limited support.
Two honest caveats. First, Soofi S is a text-only model — it writes, reasons, translates, and codes; it does not generate images, video, or audio. Second, the initial release is a base model published on Hugging Face and positioned as a foundation for further post-training (instruction tuning, preference tuning, domain adaptation) and research, not a polished end-user chat assistant out of the box. The license had not been fully finalized at release (a small share of the training data — about 1.3%, from commercially licensed Genios newspaper archives — keeps it from meeting the strictest European open-data definitions), so check the current terms on the model card before you rely on it commercially.
Soofi S is unusually good at one thing most models fumble: German that reads like a native wrote it, not a translation of an English draft. If your audience is German-speaking — or you run a German brand and refuse to sound like machine-translated English — it's a strong open drafting brain, and because the weights are downloadable you can run it on your own servers so nothing leaves the building. What it can't do is turn any of that text into a post. It has no image, video, captioning, design, or publishing layer; it's text-only, and the first release is a base model aimed at research and further fine-tuning.
That's exactly the seam Kompozy closes. Feed a Soofi S–drafted German hook, script, or angle into Kompozy and it generates the media the model can't: Persona Shorts and HeyGen avatar video with burned-in German captions, Photo Posts and face-locked Persona Photos, multi-slide Carousels and Quote Graphics rendered pixel-exact through HyperFrames, plus full Blog Articles and Email Newsletters. Then it publishes — fanning each piece across the nine supported social platforms (Instagram, Facebook, TikTok, YouTube, LinkedIn, X, Pinterest, Threads) plus email and blog, on a schedule, on autopilot, with a per-post review pipeline. Your Persona Brief holds the whole batch to one German brand voice. Soofi S owns the sovereign, German-first words; Kompozy owns the media, the brand styling, and the publish across every channel.
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 leads fully open models on aggregate German and English benchmarks.
No. Soofi S is a text-only model — it writes, reasons, translates, and codes but does not produce images, video, or audio. To turn its text into finished posts, you pair it with a generation and publishing engine like Kompozy, which renders the media and publishes across platforms.
It was deliberately trained with a much higher share of German than most open models — rising from 7.2% of the mix in an early phase to 15.3% later, versus roughly 5% for all non-English languages combined in a common reference recipe. Its pretraining report reports the top aggregate German and English scores among fully open models, ahead of OLMo 3 32B and Apertus 70B.
The weights are published on Hugging Face, but at release the license had not been fully finalized — a small share of training data (about 1.3%, from commercially licensed newspaper archives) keeps it from meeting the strictest European open-data definitions. Check the current terms on the model card before relying on it for commercial work.
The initial release is a base model aimed at research and further post-training (instruction tuning, domain adaptation), not a polished end-user assistant. It is efficient to run — its MoE design activates only ~3.2B parameters per token — but expect to fine-tune or use a tuned variant for production drafting.