Overview
- NPCI unveiled FiMI at the India AI Impact Summit and published a technical paper on arXiv detailing data curation, training methodology and evaluation.
- The model is live nationwide through the UPI Help Assistant, handling payment queries, grievance redressal, dispute resolution and mandate lifecycle tasks.
- FiMI was developed in-house to address gaps seen with general-purpose language models in high-scale, high-trust payment environments.
- The assistant currently supports English, Hindi, Telugu and Bengali, with additional Indian languages expected within six to eight months.
- FiMI employs agentic multi-step reasoning and structured tool invocation for reliability, as NPCI experiments with Mixture-of-Experts to improve scalability and efficiency.