Cathay Financial Holdings has demonstrated that fine-tuned open-source small language models can approach proprietary AI performance in financial-services applications while potentially reducing system complexity and improving deployment control.
Cathay Financial Holdings (Cathay FHC) has demonstrated that fine-tuned open source Small Language Models (SLMs) can deliver customer-intent classification performance approaching that of leading proprietary AI models, highlighting the growing viability of open-source AI in regulated financial-services environments.
Presenting its latest research at NVIDIA GTC Taipei 2026, the company showed how open-source SLMs can be fine-tuned to better understand local financial-service contexts, industry terminology, and ambiguous customer queries. The findings suggest that smaller open-source models could serve as practical alternatives to mainstream closed-source large language models (LLMs) for enterprise deployments.
According to the study, fine-tuned SLMs achieved customer-intent classification performance close to mainstream closed-source LLMs and approached that of leading proprietary models. The results provide enterprises with a benchmark for evaluating AI training and deployment strategies.
Cathay FHC also found that fine-tuned SLMs may reduce dependence on complex prompt engineering and vector retrieval modules, potentially simplifying AI system architectures while lowering future operational and maintenance complexity.
The research further showed that combining targeted model fine-tuning with carefully designed financial-domain datasets improved model stability, inference efficiency, and deployment controllability.
The study evaluated leading models from Meta, TAIDE, TAME, NVIDIA, and OpenAI. To address privacy and governance requirements, all training relied on synthetic data rather than real customer information. Techniques such as service-function clustering, single- and multi-intent dataset design, Taiwan-context localisation, and keyword expansion strengthened the models’ understanding of financial terminology and customer intent.
Potential applications include mortgage balance enquiries, credit-card payment assistance, branch-service navigation, intelligent search, and service routing. Cathay FHC plans to further explore long-context classification, financial-document understanding, and broader AI deployments tailored to the financial sector.
