Taiwan's financial industry is developing its own AI! The FinLLM project has invested nearly 70 million yuan; here's a sneak peek at the development timeline and highlights

16 financial institutions in Taiwan are promoting the FinLLM project, investing nearly 70 million yuan to build a Taiwan-dedicated financial large language model. By internalizing local regulations, it aims to address the pain points where general AI is prone to mistakes. The first version of the bank-specific model is expected to be launched by the end of this year.

16 financial institutions collaborate to develop Taiwan’s financial AI FinLLM

With the generative AI wave sweeping the globe, general large language models often face problems such as insufficient localization when handling professional financial fields, and they also struggle to connect with Taiwan’s financial industry knowledge and regulations.

In response, the FinTech Industry Alliance announced yesterday (4/22) that it will officially push forward a project for a financial large language model (FinLLM). The project brings together 16 domestic financial institutions and combines government-industry-academia resources, including those from the National Development Council, the Digital Development Department, and the Financial Supervisory Commission.

According to reports from Economic Daily and iThome, Financial Supervisory Commission Chairperson Peng Jinlong said that the financial industry is a highly regulated sector involving a large amount of complex local regulations. Currently, most of the general-purpose large language models on the market are trained on international data; if they are applied directly, there is a high risk of making mistakes in the application of regulations.

Digital Development Department Minister Lin Yijing also noted that when general models face specific country financial professional problems, they often cite foreign laws, leading to incorrect information. Developing a model with knowledge of Taiwan’s regulations and localized understanding capabilities has become an important engineering effort to ensure risk control and compliance.

Photo source: Financial Technology Industry Alliance news photo　Digital Development Department Minister Lin Yijing speaks at the press conference for Taiwan’s financial industry AI FinLLM financial large language model

By participating in this AI infrastructure, the financial industry hopes to shift compliance management from passive review to active protection, driving a comprehensive transformation of financial services and organizational operations.

The FinTech Industry Alliance also disclosed the list of participating institutions: CTBC Financial Holding, Chunghwa Post, Taishin Shin Kong Financial Holding, E.SUN Financial Holding, Cooperative Bank, Mega Financial Holding, First Commercial Bank, Next Bank, Cathay Financial Holding, Fubon Financial Holding, Hua Nan Financial Holding, KGI Financial Holding, Chang Hwa Bank, Bank of Taiwan, Land Bank of Taiwan, and Taiwan Business Bank.

FinLLM development timeline: training in May, launch the first version by year-end

As for when the financial industry’s FinLLM will be fully developed, officials revealed that the project is scheduled to officially begin model training in May this year.

The first phase will focus on the banking sector, where regulations and data foundations are more complete. It is expected to complete the initial version of the model in the third quarter of this year, and launch the final bank-dedicated model by the end of this year. Afterwards, it will be gradually expanded to the insurance and securities sectors. This Week Magazine pointed out that the entire project is expected to cost nearly 70 million yuan.

CITIC Financial’s Chief Information Officer Jia Jingguang said that the FinLLM project will combine the Digital Development Department’s “Taiwan Sovereign AI Corpus” with regulations from the Financial Supervisory Commission to establish a legally compliant training foundation. It will be handed to the local technology team Asia-Pacific Intelligent Machines for tuning and optimization, and National Chengchi University will establish a standardized evaluation mechanism to determine whether outputs are compliant.

The goal is for the system to reach the professional level of entry-level banking staff, enabling it to handle tasks such as credit assessment and financial analysis. In the future, it will also be handed over to third parties to assist with model licensing, iteration, and the establishment of an application ecosystem.

Photo source: Financial Technology Industry Alliance news photo　Group photo of attending guests at the press conference for Taiwan’s financial industry AI FinLLM financial large language model

How is FinLLM different from current approaches?

When most banks adopt generative AI at present, they generally use a retrieval-augmented generation architecture.

Jia Jingguang said that the current approach is to build a knowledge base outside the general model, allowing the model to query data in real time and then generate answers. Although this can reduce the error rate to some extent, during the process of segmenting data for retrieval, information is easy to miss. And when the amount of knowledge increases significantly, it will face technical bottlenecks such as reduced query efficiency and unstable responses.

This time, the joint development of a dedicated FinLLM differs from the retrieval-augmented architecture used in the past in that it directly internalizes Taiwan’s financial regulations and industry knowledge into the model. The system does not need to rely on external queries; it can directly understand financial logic and generate answers, clearly improving response completeness and reasoning-analytical capabilities.

This is also an important step Taiwan’s financial industry has taken after the AI Basic Law went into effect and the Financial Supervisory Commission released guidelines for AI applications in the financial sector.

In the future, AI models used in finance are expected to adopt a hybrid approach: using locally trained internalized models as the core, supplemented by external knowledge bases to provide the latest real-time information, and safeguarding decision-making through a human-machine collaboration mode—driving an upgrade in overall quality and efficiency of financial services.

Further reading:
Central News Agency reports on NTU students’ follow-up! After creating Traditional Chinese data sets for AI involving infringement, both sides have reached a settlement

People go crazy for breeding lobster! Digital Development Department: AI agents will definitely be integrated into public services, Foxconn is interested in investing in Taiwan’s computing power