This page may contain stale information. Last updated: 2026-06-28
Definition
PRAGMA (PRe-trained Banking Foundation Model) is revolut’s family of encoder-only transformer foundation models for multi-source banking event sequences, co-developed with nvidia on H100 GPUs.
Key Points
- Scale: ~25–26M users, 24–40B events, 207B tokens, 111 countries, ~28 months history
- Architecture: Profile-state encoder + per-event encoder + history encoder; 10M–1B parameter variants (PRAGMA-S to largest)
- Training: Nebius cloud H100 clusters; PRAGMA-S converges ~2 days on 16 GPUs
- Production use: Fraud screening (sub-second latency), credit scoring, engagement, recommendations
Warning
Performance metrics (64.7% fraud recall, 130% credit PR-AUC uplift) are internal revolut benchmarks — not independently verified. AML task shows 47.1% F-0.5 drop in paper; per-user encoder not graph-aware.
Related
- revolut
- foundation-models
- banking-automation
- fintech-ai
- transaction-foundation-models
- neobank-fraud
- nvidia
- 2026-06-25-revolut-pragma-banking-foundation-model