Artificial intelligence is advancing rapidly, and every new model brings promises of efficiency and performance. The German Soofi consortium has recently released the Soofi S 30B-A3B, an open foundation model that is a hybrid between Mamba and Transformer, designed specifically for German and English. With 31.6 billion total parameters, of which only 3.2 billion are activated per token, this model marks a milestone in computational efficiency. In a world where digital sovereignty is increasingly relevant, Soofi S stands out as a robust alternative for companies looking to process natural language in bilingual contexts.
The model's architecture combines 52 layers in total: 23 Mamba-2 sequence-mixing layers, 23 granular MoE (Mixture of Experts) layers, and 6 grouped-query attention (GQA) layers. Only the latter maintain a KV cache, drastically reducing memory requirements during inference. Each MoE layer hosts 128 routed experts, activating 6 per token, plus 2 shared experts. This configuration allows Soofi S to deliver performance comparable to much larger dense models, but at a significantly lower computational cost. The design is based on the Nemotron 3 Nano reference scheme without modifications, ensuring compatibility with stacks like vLLM and facilitating production deployment.
Training was carried out in three phases, consuming approximately 26.68 trillion tokens. Phase 1 processed 20 trillion tokens with a diverse, quality-tiered mixture. Phase 2 used 6.58 trillion high-quality annealing tokens, and Phase 3 extended the context window to 1 million tokens. The percentage of German data increased from 7.2% in Phase 1 to 15.32% in Phase 2, explaining the model's excellent performance in that language. Training took place on Deutsche Telekom's industrial AI cloud in Munich, using up to 512 NVIDIA B200 GPUs between March and May 2026, consuming around 253,000 GPU hours.
In terms of performance, Soofi S outperforms other open models in English (70.1%) and German (79.1%) aggregates, according to evaluations using the lm-evaluation-harness pipeline. On specific tasks like HumanEval (73.8% pass@1) or GSM8K (86.1%), it demonstrates notable capability in both mathematical reasoning and code generation. For companies working with legal or technical documents in German, scores like GLP-DE (88.8%) and INCLUDE-DE (61.2%) are particularly relevant. This opens the door to applications in insurance policy analysis, contracts, or regulatory reports.
From a business perspective, models like Soofi S represent an opportunity to develop custom software that integrates cutting-edge artificial intelligence. At Q2BSTUDIO, as a software and technology development company, we understand that the key lies not only in the model but also in how it is integrated into existing workflows. Combining such models with personalized AI services allows organizations to automate processes, improve customer service through bilingual AI agents, and extract knowledge from large volumes of text. Moreover, Soofi S's efficiency in long contexts makes it ideal for RAG (Retrieval Augmented Generation) systems in technical support or corporate knowledge bases.
Cybersecurity also plays a fundamental role when implementing these models. As an open model, companies can host it on their own cloud infrastructure (AWS or Azure) without relying on third parties, ensuring data privacy. Q2BSTUDIO offers cybersecurity services and consulting to ensure these deployments meet the highest standards. Likewise, Soofi S's native ability to handle German and English makes it a valuable tool for companies operating in both languages, enabling the development of Business Intelligence (BI) solutions with Power BI that analyze multilingual text and provide actionable insights.
In the field of automation, AI agents based on Soofi S can handle complex tasks such as drafting emails, classifying documents, or generating technical reports. The model's flexibility, being a base model without instruction tuning, allows data teams to fine-tune it with proprietary data for specific domains. Q2BSTUDIO accompanies clients throughout the entire project lifecycle, from model selection to production deployment, including integration with AWS or Azure cloud services to scale on demand.
In summary, Soofi S 30B-A3B is not just a technical advance; it is a practical tool for companies seeking to innovate with AI in a sovereign and efficient manner. Its hybrid Mamba-Transformer MoE architecture demonstrates that it is possible to combine performance with low computational cost. With the support of technology partners like Q2BSTUDIO, organizations can transform this potential into real solutions that improve productivity, security, and decision-making.
