European Financial Firms Pursue Banking AI Through Different Models

European financial firms are applying open-weight and proprietary artificial intelligence to banking tasks, with projects ranging from bunq’s running assistant to Santander experiments and Starling’s planned production rollout.

The approaches vary by company and stage. Santander’s AI Lab evaluates open-weight models for banking work, and bunq uses them in Finn. Revolut developed PRAGMA, a proprietary foundation model with NVIDIA. Starling Bank is developing a transaction foundation model with NVIDIA and cloud provider Civo.

Santander’s Model Arena evaluates open-weight models against banking tasks. The bank has also developed core-banking intelligence using NVIDIA’s Kumo model family. In selected experiments, serving costs fell to about €0.03 per interaction from a baseline near €5; those figures describe the tests, rather than a bankwide rollout.

Santander announced a three-year collaboration framework with NVIDIA covering foundation models, AI agents, fraud and anti-money-laundering tools, real-time decision-making and simulation. The framework spans model development and financial services applications, while the cost comparison reflects a narrower set of experiments.

The Dutch fintech bunq runs its Finn assistant on open-weight models hosted by Nebius. That puts the model approach into a running assistant, while Santander’s reported cost results come from selected trials.

Revolut’s PRAGMA is a proprietary foundation model developed with NVIDIA. Its development represents a different model strategy from the open-weight systems described at Santander and bunq. Starling is developing a transaction foundation model with NVIDIA and Civo.

Starling intends to train and host its model on Civo’s U.K. cloud. Fraud detection is the first planned use, followed by Starling Assistant. Its production target is early 2027.