Du Xiaoman releases XuanYuan 4.5 Financial Enhanced Large Model, with inference cost about one-tenth that of top general-purpose models
Du Xiaoman, together with the Hong Kong University of Science and Technology and the Hong Kong Generative AI Research and Development Center (HKGAI), jointly released the XuanYuan 4.5 Financial Enhanced Large Model at the 2026 Hong Kong FinTech Week. The model applies Recursive Self-Improvement (RSI) technology to model training and agent scenario adaptation, and in the tested financial scenarios achieves stronger financial scenario capabilities at an inference cost of about one-tenth that of top general-purpose models. Based on XuanYuan, Du Xiaoman has built over a thousand agents and more than 5,000 skills, with average daily Token usage exceeding 100 billion, covering areas such as risk control, credit review, customer service, and customer operations.
Body · AI translation
On November 12, Baidu World 2026 was held in Beijing, with the conference focusing on how to enhance organizational capabilities in the AI era. At the opening ceremony, Baidu founder Robin Li judged that the next act of artificial intelligence is the "intelligent organization," and for the first time proposed the "three principles of organizational evolution": breaking down job boundaries, accumulating tacit knowledge, and designing closed-loop processes. "If artificial intelligence is the new electric power, then the intelligent organization is the new factory of the AI era, the place where productivity truly happens."
During this year's conference, Baidu released a number of achievements around the four-layer architecture of "chip, cloud, model, and agent": it officially launched the large model trading platform Ciyuansuo, released the preview version of Wenxin Large Model 5.5 Flash and two Wenxin application models; Baidu Dazi, Kuku, Baidu Yijing, and Famou, four major agents, received upgrades; Baidu AI Cloud officially released the Tianchi 1024-card super node, Kunlun Chip announced a new product plan, and Apollo Go further accelerated its global operations.
For small and medium-sized financial institutions, intelligent upgrading must both solve practical problems in business and balance technology investment and application costs. Facing challenges in professional talent, technical resources, and scenario validation, how to bring AI into business links such as risk control, credit review, and customer service at a controllable cost has become an important issue for small and medium-sized financial institutions in advancing intelligence.
Recently, Du Xiaoman, the Hong Kong University of Science and Technology, and the Hong Kong Generative AI Research and Development Center (HKGAI) jointly released the XuanYuan 4.5 financial enhanced large model at the 2026 Hong Kong Fintech Week, applying self-evolution (RSI) technology to model training and agent scenario adaptation. In the tested financial scenarios, XuanYuan 4.5 achieved stronger financial scenario capabilities at about one-tenth the inference cost of general top-tier models. For small and medium-sized financial institutions that need to continuously run agents while balancing application results and investment costs, this capability combination helps lower the cost threshold for AI application.
At present, based on "XuanYuan," Du Xiaoman has built thousands of agents and more than 5,000 skills, with average daily Token usage exceeding 100 billion. These agents have deeply entered business links such as risk control, credit review, customer service, and customer operations, improving business response speed and service efficiency by independently executing tasks, collaborating with employees, and having humans take over when necessary.
In risk management, according to Du Xiaoman's disclosure, its intelligent risk control is now fully driven by agents, risk strategies dynamically self-evolve as the environment changes, and the iteration cycle has been compressed by more than half; in its own business practice, short-term delinquency risk has dropped by about 50%, and the overall annualized credit delinquency risk has been reduced by 12%. These practices provide scenario references for small and medium-sized financial institutions to use AI to enhance risk management capabilities.
In the credit review and post-loan links, the review capability of Du Xiaoman's AI credit reviewer exceeds that of 80% of manual credit reviewers, and customer waiting time has been shortened from 10 minutes to 2 minutes; the recovery rate of the AI post-loan specialist working independently is more than 8% higher than that of traditional robots. From shortening customer waiting time to improving post-loan operation results, agents are gradually integrating into financial business processes.
In the customer service and customer operations links, agents provide new possibilities for expanding service coverage. Du Xiaoman's customer service Agent can independently handle 80% of high-frequency questions and can determine user intent and call tools to solve complex business problems, with the remaining questions handled by humans as a fallback. The WeCom Agent provides customer managers with communication strategies and script suggestions; one customer manager plus the WeCom Agent can serve up to 20,000 customers, equivalent to the workload of the original 10 customer managers. For small and medium-sized financial institutions with relatively limited human resources, this type of human-machine collaboration provides a reference path for improving customer response efficiency and expanding service scope.
Du Xiaoman stated that in the future it will rely on the XuanYuan financial large model to export AI capabilities and practical experience verified in its own business to small and medium-sized financial institutions, promote the implementation of agents in scenarios such as risk control, credit review, and customer service through productized solutions, help small and medium-sized financial institutions lower technical thresholds and trial-and-error costs, and turn "organizational evolution" from Du Xiaoman's practice into capabilities that more financial institutions can replicate and apply.