Du Xiaoman Financial, The Hong Kong University of Science and Technology (HKUST) and the Hong Kong Generative AI Research and Development Center (HKGAI) jointly launched the Xuanyuan 4.5 financial-enhanced large model at the opening of Hong Kong Fintech Week on November 2, People's Daily and China News Service reported. Official figures put a 30% effect improvement across core banking scenarios — customer acquisition, risk control and customer operations — at roughly 10% of the inference cost of general-purpose models.
Xuanyuan 4.5 ships as both a language model and a multimodal model. Guo Yike, HKUST's provost, called it the industry's first self-evolving financial-enhanced model: self-evolving RSI techniques drive both training and agent-level task adaptation, the in-house Xuanyuan-Machine system participates in agent data synthesis, supervised fine-tuning and reinforcement learning, and more than 100 billion tokens of scenario-enhanced data were built in a compliance-controlled financial agent environment.
Most of the reported results come from Du Xiaoman's own credit business. The company has built over a thousand agents and more than 5,000 skills on Xuanyuan, consuming over 100 billion tokens daily; risk control is now fully agent-driven, with risk-strategy iteration cycles cut by more than half, short-term overdue risk down about 50% and annualized credit overdue risk down 12%. In credit review, AI officers outperform 80% of human reviewers and customer wait times fell from 10 minutes to 2; post-loan AI collection agents lift recovery rates by more than 8%, and service agents independently handle 80% of high-frequency queries — one relationship manager plus a WeChat-work agent can now serve 20,000 customers.
CEO Zhu Guang said Hong Kong is Du Xiaoman's international headquarters, and the company will deepen research cooperation with HKUST and expand into international markets with HKKGAI. The partnership dates to a joint AI laboratory established in March 2025; Xuanyuan 4.5 is its latest milestone.
Financial LLM competition is shifting from general capability toward the trade-off between scenario performance and cost — Xuanyuan 4.5's coordinates are one-tenth the inference cost for a 30% effect gain. Whether that combination replicates across more institutions is the point to watch.