Cathay Financial Holdings Leverages Open-Source Small Language Models to Understand Customer Needs
Cathay Financial Holdings (Cathay FHC) is continuing to drive the adoption of generative AI in the financial services sector through its internally developed AI generative framework, GAIA, and its AI-as-a-Service (AIaaS) strategy. The initiative aims to enhance operational efficiency and deliver improved services to customers. Following last year’s demonstration of the potential of large language models (LLMs) for financial services, Cathay FHC presented its latest research findings at NVIDIA GTC Taipei 2026. The research showcased the performance of a fine-tuned, open-source small language model (SLM) capable of classifying customer needs and supporting future financial services.
The study evaluated several leading open-source models developed by Meta, TAIDE, TAME, NVIDIA, and OpenAI. Initial results indicate that within the testing framework, the fine-tuned SLM can potentially reduce reliance on complex prompt engineering and vector retrieval modules. This approach can simplify system architecture and lower future operational and maintenance complexity.
The findings also revealed that a combination of carefully designed financial sector datasets and targeted fine-tuning processes can improve model stability, inference efficiency, and control during implementation. In customer need classification tasks, the fine-tuned SLM delivered performance close to widely used closed-source LLMs, nearly matching the capabilities of leading models. These results can serve as a practical reference for companies developing AI model training and deployment strategies.
From a data governance and privacy perspective, the research utilised a synthetic data approach, ensuring no customer information was used in the model training process. Cathay FHC also applied various techniques, including service function clustering, dataset design for both single and multiple needs, localisation for the Taiwan context, and keyword expansion. These methods strengthened the model’s ability to understand local financial service contexts, industry terminology, and ambiguous customer inquiries.
Cathay FHC believes this technology can be applied to various future services, such as mortgage balance inquiries, credit card payment assistance, and branch service navigation. This development serves as a foundation for intelligent search, automated service routing, and next-generation customer interaction experiences.
On the technology architecture side, Cathay FHC integrated various NVIDIA AI tools, including NVIDIA NeMo Customiser, NVIDIA NeMo Curator, and NVIDIA TensorRT-LLM, leveraging computing resources based on the NVIDIA Hopper architecture. This infrastructure supports the entire process from data generation and model customisation to inference optimisation and experimental evaluation. By utilising the NVIDIA AI ecosystem, Cathay FHC continues to strengthen its capabilities in developing sector-specific financial models, data governance, and application validation.
In recent years, Cathay FHC has consistently expanded AI innovation across various financial service scenarios. The company has built a scalable technology foundation to support internal process optimisation, customer service enhancement, financial knowledge comprehension, and AI model governance. Amid increasingly stringent regulations, data governance requirements, and rapidly changing customer expectations, Cathay FHC reaffirmed its commitment to developing safe, robust, and compliant AI research.
Looking ahead, Cathay FHC will continue to explore long-context classification technology, advanced financial document understanding, and cross-scenario AI applications. By developing model training and implementation methods specifically designed for the financial sector, Cathay FHC aims to accelerate innovation and deliver smarter, more efficient, and customer-centric financial services.