Revolut just built what might be the most ambitious AI model purpose-built for banking. Called PRAGMA, short for PRe-trained Banking Foundation Model, it’s a series of transformer-based models developed in collaboration with Nvidia that aims to handle fraud detection, credit scoring, and other financial tasks through a single unified system rather than a patchwork of specialized tools.
The numbers behind it are striking. PRAGMA was trained on roughly 40 billion banking events sourced from approximately 25 million Revolut users spread across 111 countries, amounting to 207 billion tokens.
What PRAGMA actually does differently
Traditional fraud detection in banking relies on stacking multiple machine learning models, each trained for a narrow task. One model might flag suspicious transactions, another might assess credit risk, and a third might handle identity verification. Each requires its own feature engineering, its own training pipeline, and its own maintenance.
PRAGMA takes a different approach. It uses masked modeling techniques on tokenized sequences of user interactions, then applies that understanding across multiple tasks simultaneously.
The results on Revolut’s internal benchmarks suggest this unified approach pays off. The model demonstrated a 64.7% improvement in fraud recall and a 16.7% increase in fraud precision compared to the company’s existing specialized systems. Fraud recall measures how many actual fraud cases the model catches, while precision measures how often its fraud alerts turn out to be correct.
Credit scoring saw even more dramatic gains. PRAGMA achieved up to a 130% uplift in PR-AUC metrics, a measure that captures how well a model distinguishes between good and bad credit risks across different threshold settings.
The hardware and architecture behind it
PRAGMA runs on Nvidia’s H100 GPUs within the Nebius AI Cloud infrastructure. The model comes in multiple sizes, ranging from 10 million parameters up to 1 billion parameters, designed for different use cases.
The largest variant, PRAGMA-L at 1 billion parameters, was trained using up to 64 Nvidia H100 GPUs. The smaller variants at 10 million and 100 million parameters are specifically optimized for real-time inference, meaning they can evaluate transactions as they happen rather than flagging issues after the fact.
The research paper detailing PRAGMA was submitted on April 9, 2026, and presented at Nvidia’s GTC 2026 conference. Key contributors include members of Revolut’s Research team alongside Nvidia engineers, with particular emphasis on self-supervised training techniques that reduce the need for manually labeled data.
Why this matters beyond Revolut
Revolut reports that the approach has decreased its dependency on traditionally engineered features, those handcrafted data inputs that require domain experts to design and maintain.