HKGAI, HKUST and Du Xiaoman launch self-evolving financial AI model XuanYuan 4.5
HKGAI, HKUST and Du Xiaoman launched XuanYuan 4.5, the first self-evolving financial large language model, at Hong Kong FinTech Week. The model uses recursive self-improvement technology, showing a 30 per cent performance improvement in core financial tasks while costing only 10 per cent of general
The Hong Kong Generative AI Research and Development Center (HKGAI), Hong Kong University of Science and Technology (HKUST) and financial technology firm Du Xiaoman on 2 November 2026 launched the first self-evolving financial large language model, named XuanYuan 4.5, at Hong Kong FinTech Week. The initiative aims to support the global expansion of artificial intelligence (AI) in financial technology.
According to HKGAI, XuanYuan 4.5 leverages cutting-edge Recursive Self-Improvement (RSI) technology and has demonstrated a 30 per cent improvement in performance across core financial application scenarios, including customer acquisition, risk management and customer relationship management. Its computing cost is only 10 per cent of that of general-purpose models, the centre said.
Guang Zhu, chief executive of Du Xiaoman, said: “The three parties are jointly promoting the deep application of AI in the financial sector, assisting the intelligent upgrade of Hong Kong's financial industry. Hong Kong serves as the international headquarters of Du Xiaoman Group, and we will continue to deepen our scientific research collaboration with HKUST, joining hands with HKGAI to explore international markets.”
In Du Xiaoman's actual business operations, the large model has built over a thousand AI agents and more than 5,000 skills, with daily token consumption exceeding 100 billion. The company reported that short-term overdue risks have dropped by around 50 per cent, and the overall annualised credit overdue risk has decreased by 12 per cent. In the credit approval process, the performance of AI approvers surpasses that of 80 per cent of human approvers, and customer waiting times have been reduced from 10 minutes to 2 minutes.
Enterprise WeChat AI agents can provide relationship managers with precise communication strategies and response suggestions, doubling the service capacity of a single relationship manager to 20,000 people, a workload equivalent to ten managers under the traditional model, Du Xiaoman said.
HKGAI said that through future strategic cooperation, it and Du Xiaoman will jointly promote this financial AI technology to the international market, continuing to showcase Hong Kong's research and development strength in the field of generative artificial intelligence.