Revolut Research Locks In PRAGMA Data Advantage for Banking

Revolut launched Revolut Research on 25 August 2026 as the home for its proprietary PRAGMA foundation model, already delivering 2.3 times higher accuracy on credit default risk and 65 percent more fraud cases caught from its 80 million customers.

The unit sits inside the broader AI department and will work with academic and tech partners while keeping the core models trained on Revolut’s own global data.

The launch turns years of internal model work into a named division with a public research agenda. PRAGMA already underpins live risk, fraud and recommendation systems. The new unit’s task is to push that foundation further and share methods without opening the underlying customer corpus.

Revolut Research Centres the PRAGMA Engine

Revolut Research is the institutional base for PRAGMA, a family of transformer-based models built with Nvidia that treat a customer’s full financial journey as one continuous signal rather than isolated tasks.

The company says the model learns patterns across fraud detection, credit decisions, product recommendations and customer service from a single foundation. Early internal tests on historical data produced clear lifts over previous production systems.

That single-signal design is the practical break from older stacks. Instead of separate feature stores and scorecards for each product line, one set of embeddings carries context from a payment into a credit review or a service reply. Teams then attach lighter task heads rather than rebuilding the whole pipeline.

Head of Revolut Research Anton Repushko has already presented findings at conferences including ICML in Seoul. The division plans regular scientific publications, open-source frameworks and quarterly meet-ups at Revolut offices, plus appearances at Nvidia GTC Berlin in October and ICAIF in November.

Those venues matter for hiring and for partner scrutiny. Publishing methods while holding the training data closed lets Revolut invite external challenge on architecture without surrendering the behavioural record that drives the measured gains.

The Performance Lifts in Hard Numbers

Revolut published three headline results from PRAGMA deployments on historical data.

Nvidia’s case study rounds the fraud recall improvement at 64.7 percent and precision at 16.7 percent, matching the rounded company figures. Model development cycles shortened 3 to 5 times because teams reuse shared embeddings instead of building fresh feature pipelines for each task or market.

The fraud pair of higher recall and higher precision is the operationally useful combination. Catching more cases while raising alert precision reduces both missed losses and false positives that burn investigator time. The credit lift of 2.3 times accuracy on default risk points at cleaner underwriting on the same historical population. The 41 percent relevance gain on recommendations shows the same backbone can serve growth as well as defence.

Shorter development cycles compound those lifts. When a new market or product needs a model, engineers start from embeddings already trained on global behaviour rather than assembling features from scratch. The 3 to 5 times compression in cycle time is how the research unit expects to turn conference papers into shipped features without waiting on multi-quarter rebuilds.

Why Revolut Rejects the Bolt-On Path

Pavel Nesterov, head of AI at Revolut, framed the choice directly.

To lead the future of intelligent banking, you cannot rely on third-party blueprints. We have launched Revolut Research to institutionalise our ‘build, don’t bolt on’ philosophy. By training native foundation models on our global operational data, we are giving our engineering teams an unprecedented engine to deploy smarter features faster, eliminate systemic friction, and give our customers a safer, radically better financial experience.

Pavel Nesterov, Head of AI at Revolut

Anton Repushko added that the goal is financial intelligence at the deepest layer rather than patching narrow specialised models. Traditional banks, Revolut argues, retrofit decades-old systems with off-the-shelf wrappers and end up with operational drag plus basic chatbots.

PRAGMA itself is a family of encoder-style models ranging from tens of millions to one billion parameters. Smaller versions handle real-time fraud screening at sub-second latency; larger ones prioritise accuracy where speed is less critical.

The parameter range is the engineering answer to the bolt-on critique. A single giant model cannot meet every latency budget. By keeping a family that shares training data and architecture, Revolut can place a compact encoder in the path of a live payment and reserve the larger variants for batch risk or offline analysis. That flexibility is harder to buy as a third-party wrapper bolted onto a core ledger that was never designed for continuous behavioural signals.

How the Training Data Creates the Edge

The model was PRAGMA trained on 26 million user records spanning 111 countries, covering roughly 24 to 40 billion events and 207 billion tokens over about 28 months of history.

Training ran on Nvidia H100 GPUs hosted on Nebius AI Cloud. A 10-million-parameter variant converged in about two days on 16 GPUs; larger versions took roughly two weeks on 16 to 32 GPUs. Optimisations such as sequence packing and dynamic batching delivered 2 to 5 times higher training throughput.

Revolut now serves more than 80 million retail customers across more than 40 markets and processes billions of cross-border transactions plus nearly one billion transactions reviewed monthly by security models. That live flow continuously refreshes the foundation.

• 26 million user records in the original pre-training set
• 24-40 billion banking events tokenised
• 207 billion tokens processed
• 80 million current retail customers feeding ongoing improvement

As the dataset grows the models become smarter at fraud, risk and user needs, creating what Revolut calls a proprietary intelligence advantage that traditionally structured banks struggle to match.

The scale gap between the original 26 million records and the live base of more than 80 million customers is the flywheel in plain numbers. Pre-training fixed a multi-year snapshot; production traffic keeps writing new events. Cross-border volume and the nearly one billion monthly security reviews supply the rare and adversarial cases that narrow models trained on single-market ledgers rarely see.

Sequence packing and dynamic batching turned those GPU counts into 2 to 5 times higher throughput, which is why the research unit can refresh embeddings often enough to matter for fraud patterns that shift in weeks rather than years.

Transaction Foundation Models Spread Across Finance

Revolut is not alone. Nvidia has documented a broader industry move toward transaction foundation models across finance, where a single behavioural representation trained on proprietary data replaces task-specific pipelines.

Nubank earlier published nuFormer with a different decoder architecture. Visa, Mastercard and Stripe have their own transaction-focused efforts. The shared bet is that contextual understanding of timing, device, location and history beats isolated signals.

On X, fintech observer Linas Beliūnas noted that the next banking moat may be the proprietary model that understands customers better than anyone else, turning every transaction into smarter future products. The shift is from AI features bolted onto a bank to a bank that gets smarter with every interaction.

Revolut’s open stance on frameworks while keeping the data closed reinforces the point: methods can be shared, the unique behavioural corpus cannot.

The competitive map is therefore less about who publishes first and more about who holds the densest proprietary stream. Card networks see vast payment graphs but not the full deposit, credit and in-app behaviour of a primary banking relationship. Neobanks such as Nubank and Revolut sit closer to that full journey, which is why encoder-style and decoder-style experiments are both appearing in the same peer set.

1. Shared premise – one behavioural representation beats isolated task models
2. Divergent architectures – Revolut’s encoder family versus Nubank’s decoder-style nuFormer
3. Common constraint – the training corpus stays closed even when frameworks open
4. Market signal – Visa, Mastercard and Stripe running parallel transaction-model work

That pattern explains why Revolut Research pairs conference talks and open-source releases with a hard line on customer data. The methods invite collaboration; the corpus remains the moat Linas Beliūnas described.

Shared Embeddings Cut New Model Lead Time

The 3 to 5 times shorter development cycles rest on a simple reuse rule. Once PRAGMA embeddings encode timing, device, location and history, a team building a fresh score no longer starts with raw event tables. It freezes the backbone and trains a thin head, or applies LoRA fine-tuning that updates only 2 to 4 percent of parameters.

That mechanic turns the research unit into a delivery multiplier. Lifetime value prediction, churn forecasting and anomaly detection become days-scale experiments rather than quarter-scale rebuilds. The same path supports market expansion: a new country inherits global behavioural priors and only needs local fine-tuning instead of a full feature redesign.

Sub-second fraud screening stays on the smaller encoders so latency budgets hold. Heavier accuracy work lands on the larger members of the family up to one billion parameters. Because both ends share the original pre-training on 26 million records and the ongoing refresh from more than 80 million customers, improvements in the foundation propagate without separate rewrites.

• Freeze the backbone, train a lightweight task head
• Apply LoRA so only 2 to 4 percent of parameters change
• Keep real-time paths on compact encoders for sub-second latency
• Reserve larger variants for batch accuracy where speed is less critical

Nesterov’s “build, don’t bolt on” line is operational here. Third-party blueprints rarely expose this degree of controlled fine-tuning against a bank’s own full-journey tokens. The research division’s job is to keep the shared embeddings current so every downstream team inherits the lift.

Legacy Stacks Meet a Compounding Data Gap

Traditional banks, as Revolut frames them, still stitch off-the-shelf wrappers onto decades-old cores. The result is operational drag and chatbots that lack deep transaction context. PRAGMA’s measured gains on default risk, fraud and recommendations show what a native foundation can extract when the training set already spans 111 countries and tens of billions of events.

The gap widens with traffic. Revolut’s security models already review nearly one billion transactions a month while the retail base sits above 80 million customers across more than 40 markets. Each new event sharpens the same embeddings that feed risk, growth and service. A rival that begins with siloed scorecards and thinner real-time streams must first unify data before it can train anything comparable.

Academic and tech partnerships, quarterly office meet-ups, and talks at Nvidia GTC Berlin and ICAIF help Revolut stay current on methods. They do not transfer the behavioural corpus. Open-source frameworks and scientific publications lower the cost of imitation on architecture, yet the 207 billion tokens of proprietary history remain inside the company.

AIR, the in-app assistant that executes complex tasks in one step, already shows one customer-facing outlet for that depth, even while it remains UK-only. Multimodal inputs and continual pre-training are the next internal steps so the foundation absorbs new event types without full restarts. None of those extensions require abandoning the encoder family or the Nvidia H100 training path already proven on Nebius AI Cloud.

For incumbents the implication is pacing. A research unit announced in a single press cycle cannot compress 28 months of dense history or the live cross-border flow that follows. The flywheel Revolut describes is already turning; closing it from behind means matching both model practice and the data gravity that makes the practice valuable.

What the Lab Will Ship Next

Beyond the core model, Revolut already runs advanced security reviews on nearly one billion transactions a month and powers AIR, its in-app assistant that executes complex tasks in one step (available in the UK only for now).

Future work includes multimodal inputs, continual pre-training so the model learns from new events more continuously, and expanded uses such as lifetime value prediction, churn forecasting and anomaly detection. Teams can freeze the backbone and train lightweight heads or apply LoRA fine-tuning that updates only 2 to 4 percent of parameters, spinning up new models in days.

The dedicated AI research division Revolut Research also intends to plans to publish scientific findings and open technical frameworks while hosting community events. Open research roles are listed on the company careers page.

The data flywheel is already turning. Every new customer and every new transaction sharpens the same foundation that underpins risk, growth and product decisions. Legacy banks starting from siloed models and thinner real-time behavioural streams face a widening gap that a research unit alone cannot close overnight.

Publication cadence and open frameworks will let outsiders test the architectural claims. Hiring against the listed research roles will test whether the division can staff continual pre-training and multimodal work at the same pace the live customer base grows. The measurable test remains the one Revolut already published: higher default accuracy, more fraud caught at better precision, and recommendations that feel more relevant because the model saw the full journey first.