Idea
Open-source German-English foundation model delivering efficient, sovereign AI with superior long-context and multilingual performance.
Research Paper
Core Innovation
This paper introduces Soofi S 30B-A3B, a Mixture-of-Experts hybrid Mamba Transformer that activates only a fraction of parameters per token, maintaining near-constant inference cache size as context grows. This design achieves throughput advantages over dense models for long-context and high-concurrency scenarios while maintaining competitive performance on German and English benchmarks.
Why It Matters
Organizations require efficient, sovereign AI models that respect data privacy and support multilingual use cases, especially for German and English. Soofi S reduces inference costs and scales to long contexts, enabling practical deployment in enterprise and government settings. Its open-source nature and transparent data licensing foster trust and adoption across industries.
Market Size (TAM)
$10–20B TAM for multilingual foundation models; $2–5B SAM from European enterprises and government agencies. Driven by data sovereignty concerns and demand for efficient long-context AI.
Potential Customers & Pain Points
- European enterprises – Need sovereign AI respecting data privacy
- Government agencies – Require transparent open-source multilingual models
- AI developers – Seek efficient models for long-context applications
- Cloud providers – Need cost-effective inference at scale
Business Model
Open-source foundation model with permissive licensing; monetization via enterprise support, custom fine-tuning services, and deployment on sovereign cloud infrastructure.
Competitive Landscape
- Olmo 3 32B
- Apertus 70B
- OpenAI GPT models
- Google PaLM
- Meta LLaMA
Implementation Challenges
- Competition from large proprietary models with extensive resources
- Complexity of maintaining sovereign data compliance
- Challenges in scaling open-source community adoption
Validation Strategy
- Benchmark against leading dense and MoE models on multilingual and code tasks
- Pilot deployments with European enterprises and government agencies
- Open community engagement for feedback and iterative improvements
Research Paper Overview
A Sovereign, Open-Source Foundation Model for German and English
Summary
Soofi S 30B-A3B is an open-source Mixture-of-Experts hybrid Mamba Transformer model optimized for German and English. It activates only 3B of 30B parameters per token, enabling efficient long-context and high-concurrency inference. Trained on 27 trillion tokens with emphasis on German, it matches or outperforms comparable dense models and European sovereign baselines. Developed on German Industrial AI Cloud, it will be released with full transparency and permissive licensing.