A German research consortium coordinated by KI Bundesverband has released Soofi S 30B-A3B, an open source mixture-of-experts model (Apache 2.0) trained entirely on Deutsche Telekom’s AI cloud in Munich. The model contains 31.6 billion parameters but activates only 3.2 per token, keeping inference costs close to a 3B model. The hybrid Mamba-Transformer architecture maintains nearly constant inference cache even on long contexts (up to 256k tokens), with decisive advantages in throughput and concurrency over dense models.

Why it matters. Soofi S is the first open model deliberately designed with a high share of German in the training mix (7.2% to 15.3% in later phases, versus ~5% in mainstream models). On aggregated German-English benchmarks it beats all fully open models, including OLMo 3 32B and Apertus 70B, and dominates every German benchmark in direct comparison. It’s a practical case of digital sovereignty: a European open source alternative to US/China models, trained on European infrastructure with European data, ready for self-hosting. If you work on German content or need a competitive bilingual open model, this is the first option to try.

How to get started. The model is on Hugging Face as soofi-team/soofi-s-30b-a3b. The complete pretraining report is arXiv 2607.09424.

In detail

Context

Most open source models train on data that is almost entirely English: in reference training mixes like Nvidia Nemotron’s, all non-English languages combined account for about 5%. The result is that open models competitive in English collapse on other languages, including German, which is the native language of 100 million people and the most widely spoken in the EU. The strongest closed models (GPT, Claude, Gemini) cover German, but aren’t available for self-hosting and depend on US datacenters.

What changes

Soofi S is the first open model designed to be bilingual from pretraining, with an increasing share of German reaching 15.3% in the second phase (versus the typical 5%). Training spans roughly 27 trillion tokens across three phases: 20T of broad mix (web, code, math, domains), 6T from higher-quality sources to refine patterns, and a brief third phase to extend context to 1M tokens. German data comes from HPLT, Commons Corpus, FinePDFs, FineWiki, the commercial Genios corpus (193M articles from 916 German publications), and mechanically translated or generated text.

The architecture follows Nvidia Nemotron 3 Nano unchanged: hybrid Mamba-2 + attention mixture-of-experts, 31.6B total parameters, 3.2B active per token. Only 6 of 52 layers maintain a KV cache that grows with context; the rest use fixed-size states. The practical advantage: at 40k tokens with 32 parallel requests, Soofi S generates roughly eight times more tokens per second per GPU compared to dense 14–24B models, and throughput stays flat from 4k to 256k tokens while dense models collapse. The only model showing similar behavior in testing is Qwen3.5 35B-A3B, also hybrid.

The numbers

On aggregated English-German benchmarks, Soofi S beats all fully open models, including OLMo 3 32B (Allen Institute) and Apertus 70B (ETH/EPFL). On every German benchmark in comparison, Soofi S ranks first, sometimes with double-digit margins. On code: 73.8% HumanEval, 70.2% MBPP, 84.2% MBPP-DE (German variant), the best results among open models. On INCLUDE-DE (German regional knowledge) it ties Qwen3.5 35B-A3B at 61.2 points. Versus the Nemotron baseline, the German data recipe gains 15.1 points on language and 9.6 on GPQA-Diamond (science), without losing ground in English. The weak point is German competition math (56 points on Minerva MATH-DE), where it trails the leaders.

The implications

This is the first case of an open source model trained entirely on European cloud infrastructure (Deutsche Telekom, Munich) with data weighted for a non-English language, demonstrating competitive bilingual performance. The Apache license makes it usable in production, including commercial settings, without dependence on US or Chinese APIs. For those working on German content or seeking a self-hostable bilingual open model, it’s the first option to evaluate. For those building services on European data, it’s a signal that digital sovereignty in AI can stand on technical ground, not only political.

The limits

The comparison is with open source models: top closed models (GPT-5.6, Claude Opus 4.8) remain ahead in absolute terms. Soofi S has 31.6B total parameters but activates only 3.2B per token: its behavior is closer to a 3B model than a 30B on tasks requiring full capacity in parallel. The report doesn’t cover evaluations on safety, bias, or robustness against prompt injection. German competition math remains a known weak point.