Soofi S: Germany's Open 31.6B MoE Model Leads Benchmarks

Introduction

On July 13, 2026, a German research consortium released Soofi S, a fully open-weight language model built for industrial applications rather than as a general-purpose chatbot. The project was coordinated by the KI Bundesverband (German AI Association) and brought together Fraunhofer IAIS, Fraunhofer IIS, DFKI, several universities, and companies including Ellamind and Merantix Momentum.

Soofi S matters because it addresses two goals at once. It aims to prove that a fully open model can lead other open models on both English and German benchmarks, and it aims to reduce European dependence on US and Chinese AI labs. The release positions Soofi S within Europe's broader push for sovereign AI capacity, with strong German-language performance and an eye toward industrial use cases.

Feature Overview

Hybrid Mamba-Transformer Architecture

Soofi S uses a mixture-of-experts (MoE) design with 31.6 billion total parameters, but only about 3.2 billion parameters are active per token. This sparse activation keeps inference costs closer to those of a much smaller dense model while retaining the capacity of a larger network.

The architecture itself is a hybrid of Mamba (a state-space model family) and Transformer layers. Out of 52 total layers, only 6 maintain a key-value (KV) cache, the memory structure that traditional transformers rely on for attention over prior tokens. Because most layers avoid this cache, Soofi S can process long sequences without the steep computational cost that dense transformers incur.

The practical result is that throughput stays nearly flat from 4,000 tokens up to 256,000 tokens. Dense transformer models typically show substantial throughput degradation as context length grows, since KV cache size and attention computation scale with sequence length. Soofi S's design sidesteps much of that penalty, making it well suited to workloads that require processing long documents, logs, or codebases.

A Three-Phase, 27-Trillion-Token Training Pipeline

Soofi S was trained on roughly 27 trillion tokens across three distinct phases. Phase 1 covered about 20 trillion tokens of foundational pretraining. Phase 2 added roughly 6 trillion tokens of higher-quality refinement data. Phase 3 used approximately 188 billion tokens specifically to extend the model's usable context window up to 1 million tokens.

One notable shift across these phases is the growing share of German-language data. German content made up 7.2% of Phase 1 training data, rising to 15.3% in Phase 2. This deliberate increase in German-language representation is a direct driver of the model's strong performance on German benchmarks.

Training Infrastructure

Training ran from March through May 2026 on Deutsche Telekom's Industrial AI Cloud in Munich, using up to 512 Nvidia B200 GPUs and consuming roughly 253,000 total GPU-hours. The team used NVIDIA's open-source AI training framework for the run. The project received funding from Germany's Federal Ministry for Economic Affairs and Energy, tied to the EU's NextGenerationEU program, with roughly €20 million cited in reporting on the project.

Benchmark Results

Soofi S leads other fully open models on aggregate benchmarks in both English and German. It also scored 73.8% on HumanEval, a widely used code-generation benchmark.

According to the model's developers and independent reporting, Soofi S outperforms OLMo 3 32B and Apertus 70B on these aggregate scores. The comparison against Apertus 70B is particularly notable because that model has more than double Soofi S's parameter count, yet Soofi S still leads it on the reported aggregates. Exact benchmark figures for OLMo 3 32B and Apertus 70B were not published alongside the Soofi S results, so the comparison should be read as a reported ranking rather than a side-by-side score table.

Usability Analysis

Soofi S was released initially as a base model. Post-trained variants tuned for dialogue and agentic use were described as forthcoming at launch but were not yet available. This means the current release is aimed primarily at researchers, model developers, and companies with the capacity to fine-tune a base model for their own applications, rather than at end users looking for a ready-to-use chat assistant.

Fraunhofer IAIS researcher Nicolas Flores-Herr underscored this framing directly: "Soofi S is not intended as yet another general-purpose chatbot, but as a technical foundation for industrial AI." That statement reflects the model's intended audience: teams building industrial AI products who need a strong, efficient, long-context base model they can adapt, rather than consumers seeking an out-of-the-box assistant.

Model weights and intermediate training checkpoints are published on Hugging Face, which supports both immediate experimentation and research into training dynamics. The consortium designed the release to meet the Open Source Initiative's Open Source AI Definition, though final formal license terms were not fully finalized at the time of release. Prospective users should treat licensing as an open item to verify before committing to production use.

Pros and Cons

Pros

• Leads other fully open models on both English (70.1) and German (79.1) aggregate benchmarks
• Efficient MoE design activates only about 3.2 billion of 31.6 billion total parameters per token
• Hybrid Mamba-Transformer architecture keeps throughput nearly flat from 4K to 256K tokens
• Weights and intermediate training checkpoints are published openly on Hugging Face
• Strong German-language performance addresses a gap left by many US- and Chinese-led open models

Cons

• Released only as a base model; dialogue and agentic variants are not yet available
• Final formal license terms were not finalized at launch, leaving a legal open item
• Benchmark leadership is measured against other fully open models, not closed frontier systems from OpenAI, Anthropic, or Google
• Requires fine-tuning expertise to turn into a deployable assistant, limiting immediate usability for non-technical teams

Outlook

Soofi S is best understood as an infrastructure milestone in Europe's sovereign AI strategy rather than a finished product. Its long-context efficiency and German-language strength make it a candidate foundation for industrial applications such as document processing, logistics, and manufacturing systems that need to handle long, domain-specific inputs.

The next steps to watch are the release of post-trained dialogue and agentic variants, and the finalization of the model's license under the Open Source AI Definition. Both will determine how quickly companies can move from experimentation to production deployment. If the consortium follows through on full openness, including training code and finalized licensing, Soofi S could become a reference point for European-built, openly licensed foundation models.

Conclusion

Soofi S demonstrates that a fully open, German-led consortium can produce a model that leads other open models on aggregate benchmarks while introducing genuine architectural efficiency gains for long-context processing. It is not yet a consumer-ready assistant, and its licensing remains unsettled. For researchers, industrial AI developers, and organizations evaluating sovereign, European-built foundation models, Soofi S is a release worth tracking closely as post-trained variants and final license terms arrive.

Editor's Verdict

Soofi S: Germany's Open 31.6B MoE Model Leads Benchmarks earns a solid recommendation within the open source space.

The strongest case for paying attention is leads other fully open models on both English (70.1) and German (79.1) aggregate benchmarks, which raises the bar for what readers should now expect from peers in this space. Reinforcing that, efficient MoE design activates only ~3.2B of 31.6B total parameters per token adds practical value rather than just headline appeal. The broader signal worth registering is straightforward: soofi S activates only about 3.2 billion of its 31.6 billion total parameters per token via a mixture-of-experts design. On the other side of the ledger, released only as a base model; dialogue and agentic variants are not yet available is a real constraint, not a marketing footnote, and it should factor into any serious decision. Layered on top of that, final formal license terms were not finalized at launch, leaving licensing as an open item narrows the set of teams for whom this is an obvious yes.

For developers building locally, infrastructure engineers, and anyone preferring transparent, modifiable software, this is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.