Sevren is a European AI research lab that deconstructs and rebuilds open models to align with regional requirements, running its research on dedicated European compute. The company operates Faberon, an autonomous research system that continuously tests and refines model improvements, with reserved compute capacity also powering ongoing development. Sevren's model delivers frontier AI capability tailored to European languages, regulations, and values, with the compute allocation becoming more efficient over time.
Funding
Funding not disclosed
Founders
Product
Problem
Open AI models arrive pre-shaped by the defaults of their creators, reflecting other languages, other regulators, and other ideas of what a model should refuse. This forces European organizations to choose between adopting models misaligned with local requirements or forgoing frontier capability altogether. Additionally, running serious AI research requires sustained access to large-scale compute, which is often unavailable under terms that guarantee data stays on the continent.
Solution
Sevren takes the best available open models, takes them apart, measures what they actually do, and rebuilds the parts that matter for European users. The company operates on dedicated European compute that is committed rather than borrowed, ensuring data never leaves the continent and the research queue is under its own control. At the core of the effort is Faberon, an autonomous research system that proposes directions, writes code, runs jobs, and evaluates results against existing knowledge, with most output being failure by design. The system continuously improves the serving layer between models and hardware, meaning a reserved allocation delivers more work each year on the same bill. All gains from this research are released publicly.
Target Audience
Primary customers are European enterprises and public-sector organizations that need frontier AI capability aligned with local languages, regulations, and data-sovereignty requirements, and that require sustained, dedicated compute for their own AI research.
Features
- Faberon, an autonomous research loop that proposes, builds, measures, and severs model directions without human intervention
- Dedicated European compute infrastructure with committed, never-oversubscribed capacity that keeps data on the continent
- A judgment framework that prevents the loop from fooling itself, including rules for how directions earn larger budgets and what findings must survive before being believed
- Slack compute recycling, where unused reserved cycles automatically run research instead of being wasted
- Continuous optimization of the serving layer, making reserved allocations more efficient over time
- Public release of research gains, including improvements to model serving and capability