Soofi – Sovereign Open Source Foundation Models
Soofi S introduces a 30B Mixture-of-Experts model, trained on 27 trillion tokens, as Europe's first building block for sovereign open-source industrial AI, prioritizing German and English texts. This initiative from a German consortium seeks to establish independent, high-performance AI, attracting attention for its technical ambition and questions regarding its open-source claims. The project is currently in practical testing, inviting industry partners to explore its potential in various industrial applications.
The Lowdown
Soofi S is presented as the inaugural model in a new European family of "Sovereign Open Source Foundation Models," spearheaded by the Soofi consortium. This 30B Mixture-of-Experts (MoE) model, trained on a massive 27 trillion tokens, is primarily designed for high-performance industrial AI applications in German and English contexts. The project emphasizes transparency, adaptability, and usability on sovereign infrastructure.
- Model Specifications: Soofi S is a 30B Mixture-of-Experts model, trained with 27 trillion tokens, focusing on German and English texts.
- Target Applications: It is aimed at industrial use cases requiring control, transparency, and efficiency, including technical/regulatory document processing, code generation, agentic AI systems, and company-specific applications.
- Current Status: The model is undergoing practical testing with industry partners, with a general public release pending.
- Collaboration: Soofi is a consortium of leading German research institutions and innovative startups, working to build an independent, high-performance, and trustworthy AI for Europe.
- Engagement: The project hosts workshops (Open Source Sovereign Foundation Models - OSFM) and participates in industry events to foster collaboration and discussion on European AI ecosystems.
Soofi's vision is to provide a foundational AI model that champions European technological independence and offers robust solutions for critical industrial sectors, currently seeking partners to expand its testing phase.
The Gossip
Open Source Scrutiny
Commenters immediately questioned the project's 'open source' claims, noting that while the site mentions 'license free,' this isn't synonymous with true open source. It was clarified that training and inference code are available on GitHub, but the model weights themselves are not yet openly released, with a promise for future availability.
Technical Triumphs
One comment highlighted a linked arXiv paper, pointing to detailed technical aspects of Soofi S, particularly its 'long-context serving efficiency.' It was noted that the model shows a 'small win in the small model class,' indicating its competitive performance in specific technical benchmarks.
Sovereignty Skepticism
While the story emphasizes 'sovereign' European AI, one commenter expressed cynical views towards individuals associated with 'digital value creation' initiatives, implying a disconnect between grand claims and tangible contributions, thereby casting doubt on the practical implementation or genuine impact of such 'sovereignty' projects.