kausable raises €12M seed to build AI that adapts without retraining

kausable, a Heidelberg-based AI startup founded in 2025 by three physicists from Heidelberg University — Johannes Haux (CEO), Dr. Benjamin Herdeanu (CTO), and Gregor Ramien (COO) — has raised a €12 million seed co-led by UVC Partners and Entourage, with HTGF and Mätch VC also participating. Angel investors from Black Forest Labs, OpenAI, Google DeepMind, Noxtua, and ELLIS (the European Laboratory for Learning and Intelligent Systems) round out the cap table.

The company builds what it describes as "reasoning-first" frontier AI: models that construct internal representations of the world — world models — and use them to forecast and reason about new situations from minimal data, rather than requiring the large-scale retraining that standard AI systems depend on. The previous funding was a €1.5 million pre-seed; the €12M seed is the company's first institutional round.

The retraining problem

The standard critique of current AI is capability: systems that hallucinate, confabulate, or fail on complex reasoning chains. kausable's critique is different — it is about adaptability. Every deployed AI system exists in a world that changes. Markets shift, systems evolve, distribution drifts. The dominant response is retraining: gather new data, run new compute, update the model. That cycle is slow and expensive at scale, and it assumes that new data is available — an assumption that breaks in novel environments where by definition you don't have much of it.

kausable's architecture is built around TipPFN, a zero-shot forecasting system designed for complex dynamic systems. Zero-shot means the system makes useful predictions on distributions it has not seen before, without per-task fine-tuning. Dynamic systems — the kind where the underlying rules evolve rather than stay fixed — are exactly the regime where static models trained on historical distributions fail most visibly.

"The world evolves rapidly, and so should AI," said Johannes Haux. "However, even the most capable current AI systems are very static in how they interpret the world and need constant, time-consuming, and costly retraining."

What the syndicate signals

The UVC Partners and Entourage co-lead is notable beyond the amount. UVC is the Munich-based early-stage fund with a consistent record in European AI infrastructure; Entourage, co-founded by Pieterjan Bouten, has backed several companies from the European AI research ecosystem. The angel roster — drawn from Black Forest Labs, OpenAI, DeepMind, and ELLIS — provides research credibility that matters in a field where the boundary between frontier research and commercial AI is still being actively negotiated.

HTGF (High-Tech Gründerfonds), which also backed Pixel-Flo earlier this year, is the fourth institutional participant. The German public-private seed fund's involvement is a routine checkpoint for German deep-tech Seed rounds at this size.

The €12 million funds three things: team expansion from nine to the scale needed for enterprise sales, continued TipPFN development, and the commercial translation of the world-model architecture into deployable products. The question the capital must answer is not whether the approach is intellectually compelling — the research pedigree makes that credible — but whether it solves problems that enterprises will pay to have solved at the required precision and reliability.