PRAGMA
#4 in AI Automation & Workflowsrevolut · since 2026-04-09 · 25× · last seen Jun 30, 2026
PRAGMA (PRe-trained Banking Foundation Model) is a family of encoder-style Transformer models developed by Revolut in collaboration with NVIDIA, available in 10M, 100M, and 1B parameter configurations. It was pre-trained via self-supervised masked modelling on approximately 40 billion banking events (~207 billion tokens) from around 25–26 million users across 111 countries. PRAGMA replaces siloed, task-specific models with a single shared embedding backbone covering fraud detection, credit scoring, lifetime value prediction, product recommendation, and communication engagement. It is a proprietary, internally deployed model with no public checkpoint; its architecture is published as an arXiv preprint (2604.08649).
Features
| Compliance/Certification | Runs on European cloud infrastructure (Nebius, formerly Yandex Cloud) with focus on GDPR compliance |
| Deployment Model | On-prem/cloud hybrid: training on dedicated H100 clusters, production inference via Nebius AI Cloud/Token Factory, LoRA fine-tuning for specialization |
| Use Case Scope | Credit scoring, fraud detection, customer lifetime value prediction, communication engagement, recurring transaction detection, product recommendation |
| Integrations | Integrated into Revolut's Chat Orchestrator/AIR (AI assistant) and FinCrime agents; NVIDIA cuDF/cuDNN and Nebius Token Factory for inference |
| License | Research paper published under CC BY 4.0 on arXiv; model weights/code not publicly released |
| Platform | Trained and operated on Nebius AI Cloud with NVIDIA H100 GPUs (training on 64 H100s, production on 200+ H100s) |
| Price | No public product/no pricing – internal research/production model at Revolut, not commercially distributed as a standalone product |
| Release Date | arXiv preprint published/submitted on April 9, 2026 |