Du Xiaoman Launches Xuanyuan 4.5, a Self-Evolving Finance Model

What happened

Xuanyuan 4.5 is the first finance-enhanced model to apply self-evolving RSI technology to model training and to agent-scenario target adaptation, comprising a large language model and a multimodal model. The in-house Xuanyuan-Machine takes part in agent data synthesis, supervised fine-tuning and reinforcement learning, so the model keeps adapting to finance tasks.

According to the launch announcement, Xuanyuan 4.5 uses RSI technology to improve results by 30% in core finance scenarios such as acquisition, risk control and customer management, with inference cost at about 10% of top general-purpose models. For complex multi-step business, the model combines long-chain reinforcement learning, credit standard-operating-procedure rewards and online policy distillation, with more than 100 billion tokens of scenario-enhanced data built in a trusted and compliant finance-agent environment.

On deployment, Du Xiaoman said it has built more than 1,000 agents and over 5,000 skills, with daily token usage above 100 billion. In credit business, intelligent risk control is fully agent-driven, the risk-policy iteration cycle has been compressed by more than half, short-term overdue risk fell about 50%, and annualized credit overdue risk fell 12% overall. In credit review, AI reviewers exceed 80% of human reviewers in review capability, and customer wait time fell from 10 minutes to 2 minutes.

Key facts

1. Publishers and date: Du Xiaoman, HKUST and HKGAI, Nov 2, 2026, at the opening of Hong Kong FinTech Week
2. Technology: the first finance-enhanced model applying self-evolving RSI technology to training and agent-scenario adaptation
3. Model composition: a large language model plus a multimodal model; in-house Xuanyuan-Machine handles data synthesis, fine-tuning and reinforcement learning
4. Launch figures: core finance-scenario results up 30%, inference cost at about 10% of top general-purpose models
5. Scenario data: more than 100 billion tokens of scenario-enhanced data in a trusted, compliant finance-agent environment
6. Deployment scale: more than 1,000 agents, over 5,000 skills, daily token usage above 100 billion
7. Cooperation background: Du Xiaoman and HKUST set up a joint AI laboratory in March 2025, working with HKGAI