Taiwan’s financial sector has launched a joint AI initiative aimed at building a large language model tailored to local compliance and banking needs. The project, called FinLLM, was initiated by the Financial Technology Industry Alliance under the guidance of the Financial Supervisory Commission, with CTBC Financial Holding serving as convener. It was formally launched on July 23 at FinTechSpace.
A shift away from isolated model development
The consortium brings together 16 financial holding companies and banks, along with support from the Taiwan Academy of Banking and Finance, the Institute for Information Industry, and National Chengchi University’s FinTech Research Center. According to the project description, financial institutions previously had to gather datasets, train models, and run evaluations on their own, while also handling licensing and compliance risks separately.
The participating institutions include CTBC Financial, Cathay Financial, Fubon Financial, Taishin Shin Kong Financial, SinoPac Financial, Taiwan Cooperative Financial, Mega Financial, First Financial, Hua Nan Financial, KGI Financial, Chang Hwa Bank, Bank of Taiwan, Land Bank of Taiwan, Taiwan Business Bank, Chunghwa Post, and Next Bank. Research and technical partners such as the Taiwan Academy of Banking and Finance, the Institute for Information Industry, National Chengchi University’s FinTech Research Center, and Asia Pacific Intelligent Machine are also part of the effort.
Built on Taiwan’s sovereign AI corpus
FinLLM is not being trained from scratch. Its base layer comes from the Ministry of Digital Affairs’ sovereign AI corpus, a government-led dataset designed around Taiwan’s local language, regulations, and archival materials. The corpus has already been included in one of the national AI development action plans.
On top of that general foundation, the financial consortium plans to feed in industry-specific materials such as FSC regulations, corporate governance rules, bank operating manuals, wealth management product disclosures, and risk disclosure documents. The goal is to improve model performance in Taiwan-specific regulatory contexts where large US-developed models may return only broad or incomplete answers.
Licensed data and shared evaluation are central
The project also emphasizes legally authorized data and a shared review process among member banks to create standardized evaluation benchmarks. In the financial industry, data licensing and the legal exposure attached to it can be more difficult than computing cost itself, and that issue appears to be one of the main reasons for using a consortium structure.
Based on the published details, FinLLM is intended to connect the full process from data licensing and model training to evaluation and operational planning, with the stated aim of creating a financial language model better aligned with Taiwan’s regulatory and business environment.