SOOFI (Soofi S)

innFactory AI Consulting from Rosenheim tracks sovereign AI projects for German mid-market companies closely – as a member of the German AI Association (KI Bundesverband), which coordinates the SOOFI consortium, we follow Soofi S with particular interest. The model addresses exactly the sovereignty and GDPR requirements we work on with clients across the DACH region every day.

What is SOOFI?

SOOFI stands for Sovereign Open Source Foundation Models – a German consortium project developing sovereign, open-source AI language models “made in Germany”. The goal is an AI that is trained and operated in Europe, that reflects European values, transparency and the EU AI Act (“compliance by design”) from the ground up, and that reduces German companies’ dependence on US and Chinese models.

The project is coordinated by the German AI Association (KI Bundesverband) in Berlin and funded with around EUR 20 million by the Federal Ministry for Economic Affairs and Energy (BMWE) under the European IPCEI-CIS programme. Training runs on Deutsche Telekom’s Industrial AI Cloud in Munich – with the sovereign compute infrastructure provided by T-Systems.

Consortium partners (selection)

• Fraunhofer IAIS and Fraunhofer IIS – AI research and model training
• DFKI – German Research Center for Artificial Intelligence
• TU Darmstadt, University of Würzburg, Leibniz University Hannover (L3S), Berliner Hochschule für Technik (BHT)
• The startups Ellamind and Merantix Momentum

Soofi S: the first building block

An important framing: Soofi S is the first building block of an open European model family – not the final target model. The overarching project goal is an open LLM with around 100 billion parameters as a base for European companies. Soofi S (30B) is the first published step toward that.

Architecture: efficiency through hybrid MoE

Soofi S is a mixture-of-experts model with ~31.6 billion total and only ~3.2 billion active parameters per token (hence the “30B-A3B” label). Compute cost is therefore on the level of a 3B dense model. The architecture is a hybrid Mamba-2/Transformer design in the style of NVIDIA’s “Nemotron 3 Nano”:

• 52 layers: Mamba-2 layers, granular MoE layers with shared experts, and a few grouped-query-attention layers
• 128 routed experts, of which 6 are active per token
• Only a few layers hold a KV cache – keeping memory footprint nearly constant as the context grows

The context window was extended to 1M tokens in the final (long-context) training phase. The paper is also open about the limits of such a compact model: with only ~3 billion active parameters, capacity for factual knowledge is limited – Soofi S is optimised for efficiency and transparency, not for maximum breadth of knowledge.

The paper on the 30B model

The pretraining report is titled “A Sovereign, Open-Source Foundation Model for German and English” (Soofi S Pretraining Report v1.0, arXiv 2607.09424, submitted 10 July 2026). It is authored by the ~30-person Soofi Team, including senior researchers such as Kristian Kersting, Andreas Hotho, Wolfgang Nejdl, Alexander Löser and Joachim Köhler from Fraunhofer IAIS, TU Darmstadt, DFKI and the Lamarr Institute.

The key methodological points:

1. Full data transparency: per-source token accounting, so that roughly 99% of the training mix can be independently reconstructed.
2. Bilingual depth over broad multilinguality: the German share was deliberately up-weighted (from 7.2% to 15.32% in the annealing phase), including a dedicated German evaluation suite.
3. Deliberate architecture reuse: the Nemotron 3 Nano design was adopted unchanged to isolate the effect of the data recipe and ensure usability in existing inference stacks (e.g. vLLM).
4. 3-phase curriculum: a diverse pretraining phase (~20T tokens), high-quality annealing (~6.6T tokens) and a long-context phase.

Training data and compute

In total, around 27 trillion tokens (26.68T per the paper) were processed – from open sources such as Nemotron-CC, HPLT, German Commons, FineWiki and FinePDFs, plus the commercially licensed Genios corpus (193 million newspaper and professional-press articles). Training ran on up to 512 NVIDIA B200 GPUs (64 DGX-B200 nodes) over roughly 253,000 B200 GPU hours between March and May 2026 – on 100% renewable energy with water cooling from Munich’s Eisbach.

Benchmarks

Per the paper, Soofi S is the best fully-open model on German and English (compared with models such as Olmo 3, Apertus 70B and EuroLLM):

What stands out is the efficiency: with only ~3B active parameters, Soofi S matches dense 14–27B models and, in the paper’s throughput comparison, is around 9× faster than a dense 14B model – with nearly flat throughput from 4K to 256K context.

The weaknesses should be communicated fairly: German competition-level maths trails top open-weight models, and factual-knowledge capacity is limited by only ~3B active parameters. A sovereign model is not automatically a frontier model – critical voices rightly point out that “sovereign” does not mean “superior”.

Sovereignty, licence and availability

Soofi S was trained entirely in Germany and publishes weights, intermediate checkpoints, training and evaluation code, and a detailed data inventory – meeting the Open Source AI Definition 1.0. On the licence there is currently no final decision: sources give differing information, and because around 1.3% of the Phase 1 data (Genios; ~1% of the total corpus) is commercially licensed, the model does not fully meet the stricter European open-data requirements. We therefore phrase this deliberately carefully: open source, final licence still open.

The first weights are on Hugging Face (Soofi-Project/Soofi-S-Base). So far there is no public API and no pricing – the base model is an unaligned research and fine-tuning model, not intended for direct end-user use. The consortium is currently looking for industry partners for the next testing phase.

Assessment for DACH companies

Soofi S positions itself between the multilingual EU sovereignty projects (OpenEuroLLM, EuroLLM, Teuken) and the international top open-weight models. Per the paper, it is the first European sovereign model at the same capability-per-active-parameter frontier as the strongest international open-weight releases – with full transparency.

For productive use today, however, Soofi S is still at a research/preview stage and not a ready-to-deploy product. If you need a sovereign, GDPR-compliant LLM strategy right now, established options remain the right entry point:

• Mistral for open, EU-based LLMs at a current maturity level
• Aleph Alpha (Pharia) for German-language focus
• Qwen or GptOSS for self-hosting at high maturity

Integration with CompanyGPT

Once Soofi S provides production-ready instruct variants, it can – like other open-weights models – be integrated into a self-hosted platform such as CompanyGPT and operated GDPR-compliantly in your own EU cloud. What makes Soofi strategically interesting is the combination of genuine sovereignty, radical data transparency and German-language strength.

For an individual assessment of which sovereign model strategy fits your company, contact innFactory AI Consulting.