NPCI Launches FiMi Banking: India's First Sovereign AI Model for Retail Banking
At Global Fintech Fest 2026 in Mumbai, the National Payments Corporation of India (NPCI) unveiled FiMi Banking (Finance Model for India), described as India's first "sovereign compact model" purpose-built for retail banking tasks. Built on Google's open Gemma foundation and fine-tuned with Indian banking data, the model is designed not just to answer questions, but to let AI agents reason, use tools, and carry out actions within a bank's existing rules and systems.
⚡ Quick facts
- What it is: A compact model built on Google's open Gemma models, further trained by NPCI on Indian retail banking scenarios.
- Purpose: Task-execution AI agents that can reason and act within bank rules, not just respond to queries.
- Partner: HDFC Bank is collaborating with NPCI on further development and testing.
- Benchmarks released: IndicBank Bench (799 test cases) and Tau Agentic Banking Benchmark (1,000 tasks), both using synthetic data only, publicly available.
What is FiMi Banking?
FiMi stands for Finance Model for India. NPCI describes it as a "sovereign compact model," meaning the model and its training data stay under Indian institutional control, rather than depending entirely on foreign AI labs for critical banking infrastructure.
The model is built on Google Gemma, Google's family of open-weight AI models, with NPCI conducting additional training on Indian retail banking scenarios, Retail Banker International reported after the launch at Global Fintech Fest in September 2026. It extends the original FiMI model that NPCI unveiled in February for its payments ecosystem, which powers NPCI's UPI Help assistant. NPCI has not said which Gemma version or model size it used. Unlike standard chatbots, FiMi Banking is designed for task execution: AI agents built on top of it can reason through multi-step banking workflows, invoke tools, and carry out actions, all while operating within a bank's existing rules, compliance frameworks, and regulatory boundaries.
NPCI chairman Ajay Kumar Choudhary set out a guardrail in a keynote at the same festival: AI may recommend, but authentication and final settlement must follow deterministic, auditable rules, as MediaNama reported. In other words, a model can assist with banking decisions, but the safeguards for actual payments stay rule-based.
The HDFC Bank partnership
NPCI is building FiMi Banking in collaboration with HDFC Bank, India's largest private bank by assets. The partnership merges NPCI's technology and payments infrastructure expertise with HDFC Bank's deep domain knowledge in retail banking operations, customer workflows, and regulatory compliance.
HDFC Bank's involvement matters because it brings real production banking context that lab-built models lack. The two plan to build a series of banking models together, which suggests FiMi Banking is meant for eventual use across NPCI's partner banks and fintechs, though no deployment date has been given.
Two open benchmarks for banking AI
Alongside the model, NPCI released two open evaluation benchmarks designed to let banks, fintechs, and researchers test their own AI agents against standardized Indian retail banking scenarios:
| Benchmark | Scale | What It Tests |
|---|---|---|
| IndicBank Bench | 799 scripted multi-turn test cases across 6 banking domains | Agent safety, correct actions, and response appropriateness in Indian retail banking scenarios |
| Tau Agentic Banking Benchmark | 1,000 tasks across 50 scenario groups | Dynamic multi-turn testing via dual-control simulation (second model role-plays as the customer), built on the open-source τ²-bench framework |
Crucially, NPCI confirmed that both benchmarks use synthetic data only, and no actual customer, account, or institutional information is included. The technical report, evaluation framework, scoring methodology, and full results are all being made publicly available so developers, banks, fintech companies, and researchers can test their own banking AI agents against these standardized benchmarks.
NPCI and NVIDIA's open RL environment for banking agents
Also at GFF 2026, Business Today reported, NPCI and NVIDIA unveiled an open reinforcement learning (RL) training environment for banking AI agents. Built with NVIDIA NeMo's RL framework, it lets institutions train and benchmark agents on multi-turn Indian banking conversations using synthetic data, with performance judged on outcomes.
Together, the FiMi Banking model and the NVIDIA RL environment signal that NPCI is building out a complete Indian banking AI ecosystem, not just a single model, but the training infrastructure, evaluation benchmarks, and partner integrations needed to make banking AI agents production-ready.
Our take: how a bank team should use these benchmarks
The benchmarks are the most useful part of this launch for anyone outside NPCI, because they can be run against any agent. Here is what a dual-control test looks like in practice, as an illustration of the τ²-bench style rather than an actual NPCI test case:
- The setup: a second model plays a customer who says a ₹1,500 UPI payment failed but money left their account.
- A passing agent checks the transaction status with a tool, explains the auto-reversal timeline, and raises a dispute only if the status confirms a debit without credit.
- A failing agent promises a refund it cannot issue, asks for the customer's UPI PIN, or raises a dispute without checking.
Before putting a banking agent in front of customers, we would check five things:
- Run both benchmarks, not one. Scripted cases catch known failure patterns; dual-control tasks catch what happens when a customer goes off-script.
- Read the safety score separately. A high overall score can hide a few unsafe actions, and one unsafe action in production costs more than many polite answers.
- Add your own cases. Synthetic data will not cover your bank's products or edge cases; add tests built from anonymised real complaints.
- Keep payments rule-based. Follow NPCI's line: the agent may recommend, but authentication and settlement go through deterministic systems.
- Log every tool call. If an agent takes an action, you need an audit trail that shows why.
What it means for India's banking future
India processes more real-time digital payments than any other country, with UPI handling billions of transactions every month. The prospect of AI agents handling banking tasks, from account inquiries to transaction initiation, raises both enormous efficiency potential and serious regulatory questions.
FiMi Banking addresses this by keeping the AI model under Indian institutional control (NPCI is the body that operates UPI, IMPS, and other core payment systems), using only synthetic data for training and evaluation, and insisting that payment authentication remains deterministic rather than delegated to probabilistic AI.
The approach also positions India's banking AI stack as distinct from the consumer-facing AI models from OpenAI, Google, or Anthropic. Rather than using a general-purpose chatbot as-is, NPCI is taking an open base model and specializing it, with Indian banking rules and infrastructure at its core, and publishing the tests so others can check the results. That approach could become a template for other countries building sovereign AI in sensitive sectors.
Sources
- Retail Banker International: NPCI, HDFC tie up to launch sovereign AI retail banking model
- Business Today: NPCI, NVIDIA unveil open AI training environment for banking agents
- MediaNama: NPCI says AI should not approve UPI payments
- YourStory: NPCI unveils domain-specific AI model for India’s payments ecosystem
Frequently asked questions
Can I download FiMi Banking and use it myself?
Not as far as we could find. NPCI has published the benchmarks, evaluation framework and technical report, but has not said whether the FiMi Banking model weights will be released or where.
What is NPCI FiMi Banking?
FiMi Banking (Finance Model for India) is a sovereign compact AI model purpose-built for Indian retail banking tasks, unveiled by NPCI at Global Fintech Fest 2026. It is built on Google's open Gemma models and further trained on Indian retail banking scenarios.
What makes FiMi Banking different from a regular chatbot?
Unlike standard chatbots that only respond to queries, FiMi Banking is built for task execution, meaning AI agents can reason, use tools, and carry out actions within a bank's existing rules and systems, with long-context processing.
Who is building FiMi Banking and with which partner?
NPCI (National Payments Corporation of India) is building FiMi Banking in collaboration with HDFC Bank, which will contribute its banking domain expertise to further develop and test the model.
What benchmarks did NPCI release alongside FiMi Banking?
NPCI released two open benchmarks: IndicBank Bench (799 scripted test cases across six banking domains) and Tau Agentic Banking Benchmark (1,000 tasks across 50 scenario groups using a dual-control simulation) to let developers and researchers evaluate their own banking AI agents.
Does FiMi Banking use real customer data?
No. NPCI says the benchmarks use synthetic data only, with no actual customer, account or institutional information. The technical report, evaluation framework, and results are all publicly available.