NPCI and HDFC launch sovereign AI model built specifically for Indian retail banking

India’s National Payments Corporation of India, the organization that runs the country’s wildly popular UPI payments system, just made a significant leap beyond payments. At the Global Fintech Festival 2026, NPCI unveiled FiMI Banking, which it calls India’s first sovereign compact AI model designed specifically for retail banking.

The model was built on Google’s open Gemma family of models and trained on Indian retail banking scenarios. HDFC Bank, India’s largest private sector lender, is collaborating on the project to refine it further.

What FiMI actually does

FiMI Banking is designed for what the AI world calls “agentic tasks,” meaning it can reason through multi-step processes and take actions while staying within banking regulatory guardrails, processing long-context banking interactions with an understanding of Indian regulatory requirements baked in.

One notable design choice: the entire model was trained on synthetic data, not real customer information. That’s a deliberate move to ensure data sovereignty, a concept that’s become increasingly important as countries push back against sensitive financial data flowing through foreign AI systems trained on cross-border datasets.

The benchmarks are the real story

NPCI launched two purpose-built benchmarking tools alongside FiMI. The first is IndicBank Bench, which contains 799 scripted multi-turn test cases covering six retail banking domains. These simulate the kind of back-and-forth a customer might have with a banking agent, where context from earlier in the conversation matters for later responses.

The second benchmark, called Tau² Agentic Banking Benchmark, goes further. It includes 1,000 tasks drawn from 50 different scenarios that simulate real customer interactions.

NPCI plans to open-source both benchmarks, along with technical documentation and evaluation sets. This means developers, banks, fintechs, and researchers will be able to test their own AI agents against a common yardstick, all using synthetic data rather than proprietary customer records.

Why this matters beyond India

The sovereign aspect deserves attention. By training on synthetic data and building on open-source model architectures, NPCI avoids dependency on proprietary foreign AI systems for critical banking functions.

For HDFC Bank specifically, the collaboration is a way to stay at the frontier of banking technology without shouldering the full cost and complexity of building AI infrastructure from scratch. HDFC brings the domain knowledge, NPCI brings the scale and public infrastructure mandate.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.