Revolut published PRAGMA, a single AI foundation model trained on 24 billion banking events, and the early numbers show gains in fraud, credit, and product recommendations. The model is a structural departure from the way most banks have built AI: instead of running a separate machine-learning pipeline for each task, Revolut trained one Transformer backbone that every team reuses. The PRAGMA paper went up on arXiv in April 2026, and the company’s own infrastructure partner has published its own version of the numbers since. Behind the model sits a wider stack of more than 200 NVIDIA H100 GPUs running on the Nebius AI Cloud, and three model sizes covering everything from sub-second fraud screening to credit decisions that can take longer to reason through.
The Foundation Model Banks Have Not Built
The PRAGMA paper lays out the technical shape in plain terms. The model is a family of encoder-style Transformers, pre-trained with masked modelling on a large-scale, heterogeneous banking event corpus, and it is designed to spit out one user-level embedding that downstream teams can reuse for fraud, credit, lifetime value, communication engagement, recurrent transaction detection, and product recommendation.
Most banks have built the same problem into their AI stacks: a separate machine-learning pipeline for each task, with its own hand-crafted features, its own extract-transform-load process, and its own validation cycle. Launching a new use case or entering a new market meant repeating months of feature engineering each time. PRAGMA collapses all of that onto a single shared backbone trained once and fine-tuned for each downstream task with LoRA, a low-rank adaptation technique that updates only 2% to 4% of parameters at a time. Revolut’s own data shows model development cycles are now three to five times faster, per the NVIDIA case study on PRAGMA.
The paper’s authors include Maxim Ostroukhov, Ruslan Mikhailov, Vladimir Iashin, and Pavel Nesterov from Revolut Research, alongside Vince Mullin, Roman Y. Enzmann, Georgios Kolovos, and Jason Renders from NVIDIA. Nesterov is the executive director of AI at Revolut. Anton Repushko, the named research lead on PRAGMA, said in Nebius’ customer story that the team treated data-fetching and context-engineering work as more important in production than the model itself.
Three Encoders, One Embedding
The architecture breaks the user’s banking history into three pieces, and feeds each into a specialized encoder before fusing them. One branch handles profile state, another handles the individual event record, and a third branch handles the long-term history with inter-event timing captured by a smooth logarithmic transformation so the model can read both short-term spending bursts and slow life events from the same temporal frame.
PRAGMA comes in three sizes, and the size choice is the latency knob. The smallest, PRAGMA-S, runs at 10 million parameters and converges in about two days on 16 NVIDIA H100 GPUs, with sub-second latency for real-time fraud screening at the point of transaction. PRAGMA-M sits at 100 million parameters, and PRAGMA-L, the largest, runs to 1 billion parameters, optimized for accuracy-sensitive tasks where the model can take longer to think. Training the larger variants took roughly two weeks on 16 to 32 GPUs, with sequence packing and dynamic batching driving 2x to 5x higher training throughput than padded baselines, per the NVIDIA case study. The architecture is documented in the PRAGMA paper’s architecture and benchmarks.
| Model | Parameters | Training time | Use case |
|---|---|---|---|
| PRAGMA-S | 10 million | ~2 days on 16 GPUs | Sub-second fraud screening |
| PRAGMA-M | 100 million | ~2 weeks on 16 to 32 GPUs | Mid-tier accuracy tasks |
| PRAGMA-L | 1 billion | ~2 weeks on 16 to 32 GPUs | Accuracy-sensitive credit decisions |
What the Internal Benchmarks Show
Revolut’s published numbers on PRAGMA are relative improvements against the specific models the company was running before, a framing the arXiv paper flags explicitly: “We report only relative improvements, as absolute metrics are commercially sensitive.” The headline figures are a 64.7% improvement in fraud recall, a 16.7% lift in fraud precision, a 41% gain in product recommendation relevance, and a 2.3 times better result on credit default identification. The credit number is the one that catches the eye, because credit is where banks have the deepest regulatory exposure and the strongest commercial incentive to keep the model conservative.
The training economics moved in the same direction. Sequence packing and dynamic batching on the H100 clusters drove a 2x to 5x throughput improvement, which is what made regular backbone refreshes feasible in the first place. Nebius, Revolut’s cloud infrastructure partner, also reports 3x faster pre-training efficiency and a 21% increase in fraud detection precision on the model variants it ran.
Outside the production backbone, the paper includes one extra credit-risk gain that came from a different piece of plumbing. NVIDIA’s Llama-Nemotron-Embed-1B-v2, used to embed unstructured text fields like merchant descriptions, delivered a measured 16.1% improvement in credit risk prediction performance on its own. The lesson is that a foundation model gains a lot from the embedding stack wrapped around it, and the model is only as good as the data it can read. Revolut has not disclosed which vendors sit underneath the AIR assistant, but the company has stated the architecture uses zero data retention with third-party AI providers and biometric approval for sensitive actions.
The piece that gets less attention is how the model is being adapted to new tasks. LoRA fine-tuning updates only 2% to 4% of parameters at a time, per the NVIDIA case study, which is what makes the same backbone usable across fraud, credit, marketing, and product teams. Each new market entry or new use case, in practice, becomes a small adaptation exercise rather than a months-long feature-engineering project.
From Foundation Model to Customer Assistant
PRAGMA is the engine. AIR (AI by Revolut) is the driver. AIR is the in-app conversational assistant Revolut began rolling out to its 13 million UK customers on 9 April 2026, and it is the first customer-facing product of the company’s dedicated internal AI unit. Users open it by swiping down from the centre of the home screen, and from there they can break down spending, manage subscriptions, freeze cards, check investments, and buy an eSIM before a trip, all by chat, per Revolut’s AIR launch announcement.
The features are a useful stress test of the unified backbone. Each of those actions touches a different backend: payments, cards, investments, telecommunications, and account controls. The chat layer is the consumer face of a single stack underneath, and any model that gets a recommendation wrong across product boundaries now has a single owner.
The internal numbers on the AI assistant are sharper than the marketing suggests. Across Revolut’s chatbot arc, the share of support cases resolved without a human climbed from 17% to 80%, and the financial impact ran into double-digit millions of pounds, per Nikolay Donets, who leads machine learning engineering at Revolut, in his RAAIS 2026 talk transcript. The voice layer runs in 20 countries, handles around 25,000 calls a month, and resolves a customer problem roughly eight times faster than a human agent. Chat volume on the system has scaled to up to 1.2 million support tickets per month.
The Shared Backbone and the Compliance Layer
The piece that gets less attention is the plumbing around the model. Revolut ships more than 200 products, has handled over a trillion dollars in transactions, and has a machine learning model in the path of almost every one of them. The architecture that keeps all of that inside a regulated bank, in 40 countries, is a single AI gateway with a governance layer on top. Every generative workload routes through that gateway, with one set of risks, budgets, and rules covering the use cases rather than each model. The unit of governance is the use case, which lines up with the EU AI Act’s use-case-based view of risk, and the model is the lever inside it.
Pay-as-you-go providers run at around 98.5% uptime, Donets said, which sounds high until the hours are counted at the scale of a global bank. Revolut wires a fallback chain into every generative product, so a degraded primary model rolls traffic to the next one. The company’s rule for cost: do not default to the newest model in production; measure first and use the smallest model that clears the bar, a discipline that has cut cost by as much as eight times with no loss in quality.
No decision that can change someone’s life is made by an AI system. The critical calls still do not go to the model.
What Incumbents Are Up Against
Revolut is not the only financial firm publishing foundation models for transactional data. Stripe, Mastercard, and Visa have all shipped or announced their own, per the Linas Substack analysis of the convergence.
PRAGMA stands out in two ways: it is the only one fusing multiple event sources, transactions, app navigation, trading, and push interactions, into a single user-level embedding, and the only one built by a consumer neobank rather than a payments network. The neobanks and incumbents have also moved fast on the customer-facing layer. The same customer scale underpins the wider business case Revolut is making to public-market investors, per Revolut’s $115B valuation pitch. Starling Bank and NatWest both shipped their own assistants in March 2026, and the larger incumbents have active partnerships with frontier-model vendors.
The competitive context is one of the clearer points in the FinTech Weekly launch coverage: Revolut is not ahead of the field on the assistant front. Revolut’s advantage is the data volume underneath the assistant: more than one billion transactions monthly across more than 70 million global customers, and PRAGMA trained on 24 billion of those events from 26 million user records. No European neobank runs at comparable transactional volume, and no traditional bank has the event scale to match what PRAGMA is being fed.
- Starling Bank: in-app AI assistant, March 2026
- NatWest: agentic assistant, March 2026
- Barclays: partnership with Microsoft
- HSBC: partnership with Mistral
Where the Bet Is Most Exposed
The privacy posture is the part Revolut has been clearest about, and the part that matters most for adoption. AIR operates under a zero data retention policy with third-party AI providers, can only access information the customer already sees in the app, and requires biometric approval before any sensitive action. The arXiv paper goes further on the pre-training side: all data used in the work is fully anonymised, contains no personally identifiable information, and is de-identified through standard eligibility criteria before the model ever sees it. Even so, the same paper reports only relative improvements and treats absolute metrics as commercially sensitive. None of those choices eliminate the regulatory exposure that comes with one foundation model feeding fraud, credit, and product decisions for 70 million customers.
Italy’s antitrust regulator fined Revolut €11.5 million in 2026 for consumer-protection violations, an unrelated case that underlines how thin the room is for any digital bank to misstep. The wider test for PRAGMA is whether regulators in Revolut’s biggest growth markets will let a single behavioral backbone touch credit decisions, AML alerts, and product targeting at the scale Revolut wants. Revolut’s data and revenue scale give the model room that the typical challenger bank does not have, and Revolut has said it plans to expand the model into churn forecasting and anomaly detection.
- 24 billion banking events in PRAGMA’s pre-training corpus
- 26 million user records spanning 111 countries
- 1 billion transactions processed by Revolut each month
- 70 million customers across 40 markets
- 80% of Revolut support chats resolved without a human agent
Frequently Asked Questions
What is Revolut’s PRAGMA?
PRAGMA is Revolut’s proprietary foundation model for banking event sequences, an encoder-style Transformer that the company pre-trained on roughly 24 billion anonymised banking events from 26 million user records across 111 countries. The model is designed to output a single user-level embedding that downstream teams can reuse for fraud detection, credit scoring, lifetime value prediction, communication engagement, recurrent transaction detection, and product recommendation, replacing the per-task pipelines most banks run separately.
How was PRAGMA trained?
PRAGMA was pre-trained on NVIDIA H100 GPU clusters running on the Nebius AI Cloud, with the smallest 10-million-parameter variant converging in about two days on 16 GPUs and the larger 100-million and 1-billion-parameter variants taking roughly two weeks on 16 to 32 GPUs. Training used masked modelling with a key-value-time tokenisation scheme, sequence packing, and dynamic batching, with 2x to 5x higher training throughput than padded baselines. Revolut applies LoRA fine-tuning for downstream tasks, updating only 2% to 4% of parameters at a time.
How does PRAGMA compare with other banking foundation models?
PRAGMA sits alongside foundation models from Stripe, Mastercard, and Visa in what the Linas Substack analysis calls a new competitive layer in financial services. PRAGMA is the only one fusing multiple event sources, transactions, app navigation, trading, and push interactions, into a single user-level embedding, and the only one built by a consumer neobank. Its reported relative gains over Revolut’s own prior production model include a 64.7% lift in fraud recall, 2.3x better credit default accuracy, and 41% more relevant product recommendations.
What is AIR (AI by Revolut)?
AIR is the in-app AI assistant Revolut began rolling out to its 13 million UK customers on 9 April 2026. It is the first product from Revolut’s dedicated internal AI unit and operates as a conversational layer over the same stack that protects transactions. AIR runs under a zero data retention policy with third-party AI providers, requires biometric approval for sensitive actions, and lets users manage spending, subscriptions, card controls, investments, and travel essentials by chat.
Where does the data come from?
PRAGMA was trained on pseudonymised banking event histories aggregated from Revolut’s own user base: 26 million user records spanning 111 countries and roughly 24 billion events covering transactions, in-app navigation, trading, and communications over about 28 months of history. All data is fully anonymised and de-identified, with no personally identifiable information included, and the paper notes that absolute metrics are commercially sensitive, so only relative improvements are reported.
