Revolut built its own AI foundation model, trained entirely on proprietary banking data. The model is called PRAGMA, short for PRe-trained Banking Foundation Model. It’s an encoder-only Transformer architecture trained on 24 billion banking events from roughly 26 million users across 111 countries.
The model family spans between 10 million and 1 billion parameters. PRAGMA was pre-trained on 207 billion tokens derived from those 24 billion banking events, with the training process spanning 25 months from 2023 to 2025. The model uses a custom tokenization strategy designed specifically for financial records, paired with a masked modeling objective that teaches it to predict missing pieces of banking event sequences.
Revolut reports a 130% improvement in PR-AUC for credit scoring tasks compared to their prior machine learning standards. Fraud detection recall jumped 65%.
The project was a collaboration between the Revolut Research team and NVIDIA, which provided the accelerated computing infrastructure needed to train models at this scale. The foundational research was published as an arXiv preprint on April 9, 2026, with more detailed analysis released on May 3, 2026.
One of the key technical innovations is how PRAGMA generates reusable embeddings. Rather than building separate models for each banking task, the foundation model creates general-purpose representations of users and their financial behavior. These embeddings can then be fine-tuned for specific applications like credit scoring, fraud detection, or potentially customer segmentation and risk management.
PRAGMA represents one of the first large-scale attempts by a major financial institution to build a foundation model directly on proprietary banking data. Most banks and fintechs still rely on third-party AI tools or traditional machine learning pipelines built feature by feature.