FinStep Launches Financial Agent Suite, Partners With Zhongtai Securities and Shanghai Jiao Tong University

At the 2026 Inclusion·Bund Summit, FinStep, the financial vertical large-model company incubated by Jiemian Caijing and StepFun, unveiled a financial agent product matrix, launched a joint multi-agent research platform with Zhongtai Securities, and open-sourced the Alpha-R1 financial reasoning model with Shanghai Jiao Tong University, according to Jiemian News. The company said its new product lineup includes the personal finance news assistant "Xiao Caishen," the scenario-based tool CashClaw, and Finwork, an assistant for financial institutions. FinStep CEO Jia Baolong said financial agents are moving from personal assistants toward institution-level production tools, and that the company first validated "trusted sources" and "controllable output" in consumer products before extending those capabilities to institutional clients. FinStep said its cooperation with Zhongtai Securities began in July, when Zhongtai's technology R&D department launched a project on multi-agent applications in investment research. The two sides built the "Xingxuan" agent platform, which embeds FinStep's consumer and institutional tools into Zhongtai's research workflow and connects with Zhongtai's existing systems through a Skill Hub library. FinStep also said the platform will support technical development, scenario deployment, commercialization and joint talent training. On the technical side, FinStep and a team led by Associate Professor Hua Cheng of Shanghai Jiao Tong University's Antai College of Economics and Management open-sourced Alpha-R1, a dynamic asset-allocation reasoning framework based on reinforcement-learning large-model reasoning. The company said out-of-sample tests showed a Sharpe ratio of 2.23 and maximum drawdown of 6.58% on the CSI 300, and a risk-adjusted return of 2.80 with simulated annualized return of 73.52% on the CSI 1000, outperforming traditional statistical, machine-learning and general reasoning models.