Guided by the Financial Supervisory Commission and convened by CTBC Financial Holding, 16 benchmark financial holding companies and banks in Taiwan jointly launched the “Financial Large Language Model FinLLM” project, creating localized financial AI based on the Sovereign AI Corpus developed by the Ministry of Digital Affairs.
(Background: The Taiwanese government is actively laying out “Sovereign AI.” National Science and Technology Council Chair Wu Cheng-wen said: “Forming a National Artificial Intelligence Strategy Special Committee.”)
(Additional background: Taiwan’s 11 leading insurance companies have kicked off the “Claims Settlement Blockchain Alliance”! With one subscription, customers can receive claims from multiple insurers, and the application process is much easier.)

Together, the 16 financial holding companies, 2 state-owned banks, plus the Financial Training Institute, Institute for Information Industry, and National Chengchi University’s Financial Technology Research Center, are working to achieve one goal: Taiwan should have its own financial large language model.

This project is called FinLLM. It was initiated by the Financial Technology Industry Alliance formed under the guidance of the Financial Supervisory Commission, with CTBC Financial Holding serving as the convener. It was formally launched today (23rd) at FinTechSpace, the Financial Technology Innovation Park.

From “independent R&D” to “collective collaboration”

In the past, if a financial holding company wanted to develop AI, it had to collect data, train models, and evaluate them on its own. With high computing costs, scattered data licensing, and legal compliance risks shouldered individually, the result was that each company only managed to do so-so—burning a lot of money, yet with shockingly few use cases that could be implemented.

This time, the structure is different. The 16 participating institutions include CTBC Financial Holding, Cathay Financial Holding, Fubon Financial Holding, Taishin Financial Holding, E.SUN Financial Holding, E.SUN Financial Holding, Mega Financial Holding, First Financial Holding, Hua Nan Financial Holding, KGI Bank, Changhua Bank, Bank of Taiwan, Land Bank, Taiwan Cooperative Bank, Chunghwa Post, and Next Bank—virtually pulling Taiwan’s major financial players into a single project at once.

In addition, with research and technical units such as the Financial Training Institute, Institute for Information Industry, National Chengchi University’s Financial Technology Research Center, and Asia-Pacific Intelligent Machines, FinLLM plans to connect the entire chain—from data licensing, to model training and evaluation, to business-model operations and maintenance.

More importantly, the legal risks of data licensing can also be shared collectively, which for the financial industry is often more troublesome than computing costs.

The foundation is the Sovereign AI Corpus from the Ministry of Digital Affairs

FinLLM is not a model trained from scratch. Its foundation is the Sovereign AI Corpus being built by the Ministry of Digital Affairs.

Sovereign AI Corpus. Put simply, it is a national-led effort to collect Taiwan-based language, regulations, and document data as the training basis for domestic AI models. This project has been included in one of the national-level “Ten Major AI Infrastructure Initiatives,” so its rank is by no means low.

The role of the financial alliance is to further feed this general foundation with finance-specific data: regulations of Taiwan’s Financial Supervisory Commission, corporate governance codes, each bank’s operating manuals, personal wealth management product description documents, and risk disclosure files.

No matter how strong a US-based large model may be, it may still be unable to fully answer Taiwan’s financial regulatory context. For example, if you ask GPT to interpret a provision of Taiwan’s Financial Consumer Protection Act, it may provide a general explanation. But when it encounters Taiwan-specific concepts such as “offshore structured products” or “suitability review for elderly customers,” the accuracy of its answers may drop.

Through training using legally authorized data, and with the joint deliberation of alliance banks, the project establishes standardized evaluation criteria.

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.