Ant Group's Bailing Releases Finance-Enhanced Open Model for Investment Research Workflows

The model has been open-sourced. In parallel, Bailing launched FinFIRST, an open evaluation benchmark for financial search agents. Ling-3.0-flash-Fin initially targets investment research, covering information retrieval, research reasoning, valuation modeling and report drafting, with the goal of creating an end-to-end task chain from finding materials and conducting analysis to producing research output.

The financial sector is regarded as a demanding testbed for professional AI. A typical research report requires scanning annual reports, financial statements, announcements, regulatory filings and research materials, checking different reporting periods and accounting standards, performing calculations and valuation modeling, and finally producing results that professionals can review and reuse. Traditional financial large models mainly handle single-point tasks such as question answering and summarization; Ling-3.0-flash-Fin instead aims at the full workflow.

Built on Bailing's Ling-3.0-flash, the model maintains 124 billion total parameters with 5.1 billion active parameters and a 256K-long context. Through continued pre-training on financial corpora, domain post-training and tool-use optimization, the model strengthens its ability to handle complex content such as annual reports, financial workbooks and multiple research documents. It emphasizes long-horizon agentic work, linking information retrieval, evidence review, calculation, modeling and report preparation into a complete financial research task.

The model highlights four capabilities: information retrieval, research reasoning, valuation modeling and report drafting. Information retrieval prioritizes official, first-hand and highly credible sources, with attention to timeliness and statistical calibers. Research reasoning uses multi-step calculations and cross-validation to distinguish facts from assumptions and build an auditable evidence chain. Valuation modeling works within real spreadsheets, supporting formula switching, cross-sheet dependencies and anomaly troubleshooting, so AI-generated results can continue into professionals' existing workflows. The model then organizes facts, data and judgments into editable, reviewable research materials. These capabilities correspond to a complete chain from finding and verifying information to calculating, modeling and forming research output, shifting financial AI competition from single-point capabilities to full workflow capability.

FinFIRST, developed by Ant Group with professional support from CICC's investment banking team and input from more than 50 financial professionals, aims to assess how a financial agent handles information retrieval, source selection, statistical-caliber judgment and reasoning processes. In this framework, a capable financial AI must not only give a correct answer but also explain where the information comes from, what caliber is used and how the conclusion is reached.

FinFIRST covers mainland China, the United States, Hong Kong and other markets, with Chinese questions accounting for 60.2% and English questions 39.8%. More than four-fifths of the questions contain multiple related sub-questions, 61.8% require explicit calculation, 32.5% require multiple sources, and 75.6% do not specify sources, requiring the agent to search and judge autonomously. All questions use publicly accessible sources.

Ling-3.0-flash-Fin has been tested on FinFIRST, FinSearchComp Verified, Finance Agent, APEX-Agents, SpreadsheetBench and τ³-Banking, covering financial information retrieval, investment research, long-horizon task execution, valuation modeling and banking. According to the report, the model outperformed some larger flagship models overall and is competitive among same-size models.

In addition to the model, Bailing is opening up its model, tools and evaluation system. The model weights are available, and an API is open to developers, while FinFIRST is also released to lower the barrier for financial AI research and application development. This open approach allows financial institutions, research teams and developers to participate in iteration.

Finance is one entry point for Bailing's exploration of real-world AI. As AI moves from chat and search to agentic workflows, the value of models is shifting from how much content they generate to how much work they can take on. In professional fields, whether AI becomes productive depends on its ability to understand industry rules, call professional tools and be embedded into the full process of a job. From "answering a question" to "completing a job," the release reflects the next step in moving large models from capability to productivity. Ant Group's Bailing said it will continue to explore finance enhancements on larger model bases, improve long-horizon complex task handling, and expand from investment research to more financial business scenarios.