On November 2, at the opening ceremony of the 2026 Hong Kong Fintech Week, Du Xiaoman, together with The Hong Kong University of Science and Technology and the Hong Kong Generative AI Research and Development Center (HKGAI), officially released the all-new Xuanyuan 4.5 financial-enhanced large model. As an important innovation in the financial vertical domain, the model, with its innovative RSI technology, achieves a 30% improvement in effectiveness in core financial scenarios such as customer acquisition, risk control, and customer operations, while its inference cost is only 10% of that of general-purpose models.
Deep industry-academia-research collaboration: building a new high ground for internationalized financial intelligence
At the launch event, Du Xiaoman CEO Zhu Guang and HKUST Provost Guo Yike jointly lit up Xuanyuan 4.5. Zhu Guang said that HKUST and HKGAI have profound expertise in artificial intelligence and interdisciplinary research, forming a perfect complement to Du Xiaoman's consistent solid financial business experience, real application scenarios, and engineering implementation capabilities. The strong alliance of the three parties aims to promote the deep application of AI in financial vertical scenarios and help Hong Kong's financial industry accelerate its intelligent upgrade. Meanwhile, Hong Kong, as Du Xiaoman's international headquarters, will continue to deepen research collaboration with HKUST and HKGAI in the future, jointly exploring vast international markets.
Guo Yike pointed out that HKUST has always been committed to bringing cutting-edge AI research to real industry scenarios. The jointly released Xuanyuan 4.5 is the industry's first self-evolving financial-enhanced large model, and its official deployment will further consolidate and enhance Hong Kong's technological competitiveness as an international financial center.
Hardcore technical breakdown: building a self-evolving AI-native experience of "one person, a thousand faces"
As the first financial-enhanced large model in the financial field to apply self-evolving RSI technology to model training and agent scenario goal adaptation, Xuanyuan 4.5 comprises two major parts: a large language model and a multimodal model.
In model training and iteration, the in-house Xuanyuan-Machine was deeply involved in the full workflow of agent data synthesis, supervised fine-tuning, and reinforcement learning, enabling the model to continuously adapt to complex financial tasks. For complex business scenarios requiring multi-step decomposition and analysis, "Xuanyuan" integrates technologies such as long-chain reinforcement learning, credit standard operating procedure rewards, and online policy distillation, pooling the capabilities of experts across multiple fields, and has built a scenario-enhanced dataset of over 100 billion Tokens within a trustworthy and compliant financial agent environment.
In terms of personalized services, based on precise insight into financial expertise and users' deep needs, "Xuanyuan 4.5" can provide truly "one person, a thousand faces" customized solutions targeting the operational rhythms and funding needs of different customers at different development stages. For example, for customers who need short-term capital turnover and rely on receivables for cash flow, the model can intelligently match a flexible solution of "paying interest first during the loan period and repaying the principal after the payments are collected", achieving forward-looking insight into and precise matching of customer needs.
Comprehensive agent-driven: risk control, credit review, and customer service see full-chain efficiency leaps
Backed by the powerful foundation of the Xuanyuan large model, Du Xiaoman has now built more than 1,000 agents and over 5,000 skills, with daily Token usage exceeding 100 billion. Across various specific business scenarios, AI has demonstrated transformative implementation results:
Intelligent risk control: Du Xiaoman's credit risk control is now fully agent-driven with dynamic self-evolving environments. The iteration cycle of risk strategies has been compressed by more than half, and the response speed to risk changes has greatly accelerated. Among these, short-term overdue risk has dropped by about 50%, and annualized credit overdue risk has been reduced by 12% overall.
Credit review and post-loan management: In the credit review process, the review capability of AI credit reviewers has surpassed 80% of human credit reviewers, and customer waiting time has been sharply reduced from the previous 10 minutes to 2 minutes. In post-loan scenarios, the collection rate of AI post-loan specialists working independently has increased by more than 8% compared with traditional robots.
Intelligent customer service and WeCom collaboration: The customer service Agent can not only accurately judge user intent but also autonomously query the customer service knowledge base and call tools multiple times to solve complex business, just like a human employee. Currently, the Agent can independently handle 80% of high-frequency questions, with complex questions handled by humans as a fallback. In the WeCom scenario, the WeCom Agent can provide relationship managers in real time with communication strategies and script suggestions based on historical conversation records, enabling the service scale of "one relationship manager + WeCom Agent" to soar to 20,000 people, with a single person's workload equivalent to that of the previous 10 relationship managers.
The official release of Xuanyuan 4.5 continues the deep industry-academia-research collaboration achievements since Du Xiaoman and The Hong Kong University of Science and Technology established the "HKUST–Du Xiaoman Artificial Intelligence Joint Laboratory" in March 2025. In the future, Du Xiaoman will continue to join hands with HKGAI to accelerate the international expansion of financial-enhanced large model technology, contributing more Chinese innovative solutions to the digital-intelligent transformation of the global financial industry.