Inside Revolut's PRAGMA: The Foundation Model Trained on 40 Billion Banking Events ðŸ§
Architecture, performance benchmarks vs. Stripe, Mastercard, and Visa, regulatory risks, and why PRAGMA may be the most consequential AI bet in consumer finance.
👋 Hey, Linas here! Every day, I break down 3 stories shaping the future of FinTech & Artificial Intelligence - plus the money movements and trends worth tracking. First time here? 370k+ FinTech and AI leaders get this daily. Join them:
Shortly after announcing their phenomenal 2025 financial results, Revolut has published PRAGMA (arXiv: 2604.08649), a family of encoder-style Transformer foundation models pre-trained on banking event sequences from roughly 25 million users across 111 countries. The training corpus spans approximately 40 billion events and 207 billion tokens.
What’s interesting here is that the architecture replaces the old paradigm of siloed, task-specific models with hand-crafted features: a single shared backbone transfers learned representations across credit scoring, fraud detection, lifetime value prediction, communication engagement, and more.
The results are quite striking - a +130.2% lift in Credit Scoring PR-AUC, +64.7% improvement in Fraud Recall, and +79.4% gain in Communication Engagement PR-AUC over production baselines.
But the performance numbers are not the story here.
PRAGMA is a structural bet on a specific theory of value creation: that rich, longitudinal behavioral data from a Super App generates embeddings universally better than anything hand-crafted per task, and that the competitive moat in AI is no longer algorithms but proprietary event scale.
Within the past 12 months, Stripe, Mastercard MA 0.00%↑, Visa V 0.00%↑, and now Revolut have all published or announced foundation models for financial data. That convergence is definitely not a coincidence. It marks the beginning of a new competitive layer in financial services - one where the quality of behavioral representation matters more than the sophistication of any individual model.
PRAGMA is arguably the most ambitious of the group: it is the only model that fuses multiple event sources (transactions, app navigation, trading, push notification interactions) into a single user-level embedding, and the only one built by a consumer neobank rather than a payments network.
For anyone building, investing in, or competing with modern financial infrastructure, what matters now is understanding:
How PRAGMA’s three-encoder architecture actually works, and where it breaks down
Why it delivers a 130% improvement in credit scoring but a 47% degradation in anti-money laundering (AML)
How it compares, on precise technical terms, to Stripe’s PFM, Mastercard’s LTM, and Visa’s TransactionGPT
Which regulatory constraints could keep PRAGMA’s best results out of Revolut’s highest-value markets for years
Whether traditional banks can replicate any of this, or whether their organizational structure makes it structurally impossible
What the minimum data scale threshold is for fintechs and neobanks that want to build their own foundation models
Where the commercial opportunities are for startups - from thin-file credit scoring to explainability tooling
How the foundation model arms race between Stripe, Mastercard, Visa, and Revolut reshapes the competitive landscape for the entire industry
This deep dive works through each of those in detail.