kausable Raises €12M to Build AI That Adapts Without Retraining

Most AI systems become less reliable when the world stops resembling their training data.

A machine-learning model built for one factory, market or energy network may perform well until equipment changes, demand shifts or a previously unseen event occurs. Adapting it often means collecting more data, retraining the model and validating it again.

Heidelberg-based kausable has raised €12 million in Seed funding to pursue a different architecture: AI that develops causal intuitions and adjusts to unfamiliar conditions from only a small number of examples.

UVC Partners and Entourage led the round. Existing investors HTGF and Mätch VC also participated, alongside private investors connected to organizations including Black Forest Labs, OpenAI, Google DeepMind and ELLIS.

Founded in 2025 by physicists Johannes Haux, Dr. Benjamin Herdeanu and Gregor Ramien, kausable emerged from research linked to Heidelberg University and experience across AI, cybersecurity and financial services. Before the Seed round, the company raised approximately €1.5 million in pre-seed financing.

Training on possible worlds instead of waiting for real failures

Real-world data is plentiful for ordinary events. It is often scarce for the events organizations most need to predict.

An industrial breakdown, grid instability or ecological collapse may occur too rarely to provide a conventional model with enough examples. Waiting to collect more failure data is not a practical solution.

kausable trains its models on large collections of synthetic dynamic systems. These simulations expose the AI to different causal structures, noise patterns and transition scenarios before it encounters a real deployment.

The objective is not to memorize every possible system. It is to learn reusable principles about how systems change, then apply those principles when only limited real-world observations are available.

TipPFN tests the idea on events that rarely provide a second chance

The company’s first published research model, TipPFN, is designed to estimate how close a dynamic system is to a critical transition.

The model was trained entirely on synthetic scenarios and evaluated across more than 20 synthetic, simulated and real-world datasets. According to the research paper, it detected previously unseen tipping regimes without system-specific retraining and could incorporate additional context when it became available.

Examples in the study included warning of instability in an electrical grid and estimating an approaching collapse in a cyanobacteria population. The findings remain research results rather than evidence of broad commercial deployment, but they illustrate the type of problem kausable is targeting: systems where early intervention matters and historical examples are limited.

A smaller model could carry a larger economic advantage

Frontier AI development is commonly associated with expanding model size, training datasets and computing budgets.

kausable is making a different bet. If a model can adapt through in-context reasoning rather than repeated retraining, organizations may be able to deploy AI across more systems without building a separate model for every site, machine or operating condition.

That could be relevant in robotics, energy, finance, healthcare and industrial forecasting—domains where conditions change quickly and mistakes can carry material consequences. The company ultimately wants its causal reasoning layer to work alongside language, vision and other specialized models rather than replace them.

The next proof must happen outside the research environment

kausable will use the funding to expand its nine-person team and continue developing its rapid-learning frontier model.

Its central claim is ambitious: that AI can form useful causal expectations, transfer them to unseen systems and adapt without the expensive retraining cycle that defines much of modern machine learning.

Research benchmarks can demonstrate generalization under controlled evaluation. Industrial customers will demand something harder—reliable predictions from incomplete, noisy and continuously changing operational data.

kausable’s opportunity rests on closing that gap. The company does not need to prove that one model can know everything. It needs to show that a model can enter an unfamiliar system, learn quickly and become useful before the conditions change again.

Source : kausable Official Announcement