Revolut Launches AI Research Division Built Around Pragma, a Proprietary Foundation Model
Revolut Launches AI Research Division Built Around Pragma, a Proprietary Foundation Model #
Revolut has established a dedicated artificial intelligence research division, Revolut Research, centered on Pragma, a proprietary foundation model developed in collaboration with NVIDIA. The company says the model is intended to serve as a unified intelligence layer across its product stack rather than a collection of task-specific models.
Pragma is a family of encoder-style transformer models pre-trained on approximately 24 billion banking events drawn from about 26 million users across more than 100 countries. Rather than running separate systems for fraud detection, credit scoring, customer engagement, and product recommendations, Revolut has consolidated onto a single shared backbone. The backbone learns behavioral representations from raw financial event streams, covering transactions, app interactions, currency exchanges, bill payments, and support requests, and fine-tunes those representations for individual downstream applications.
According to NVIDIA’s published case study on the collaboration, the approach produced a 130% improvement in PR-AUC on credit scoring and a 64.7% increase in fraud recall over Revolut’s own production baselines. Development timelines shrank by a factor of three to five, as feature engineering, previously a process measured in weeks or months, was largely eliminated.
The research paper describing Pragma was submitted to arXiv in April 2026 and presented at NVIDIA’s GTC 2026 conference by a joint team of Revolut researchers and NVIDIA engineers. The model uses a self-supervised masked-modelling training objective adapted for discrete, variable-length banking sequences and was trained on NVIDIA H100 GPUs via Nebius AI Cloud, where sequence packing and dynamic batching made large-scale pre-training feasible within fixed compute windows. The architecture scales from 10 million to one billion parameters, and fine-tuning a new task is designed to require updating only a small fraction of model weights.
Revolut says consolidating from multiple siloed systems to a single backbone also reduces ongoing maintenance overhead and creates consistency in how teams across the business develop and iterate on AI capabilities.
Revolut says it views proprietary AI infrastructure, rather than licensed or off-the-shelf models, as a long-term competitive advantage. By owning the full stack from pre-training data through embeddings to fine-tuned applications, the company argues it can develop a more granular understanding of financial behavior than general-purpose language models allow.
Copy link