Ant open-sources finance AI model for research workflows

Ant Group has open-sourced Ling-3.0-flash-Fin, a financial model designed for real-world workflows, adding a finance-focused model to its broader Ling 3.0 artificial intelligence lineup.

The model was developed for investment research and financial analysis, where users often need traceable sources, consistent definitions and auditable outputs. According to Ant, the system was co-developed with financial institutions and industry experts, whose input shaped task design, data systems and evaluation methods.

Ling-3.0-flash-Fin uses a Mixture-of-Experts architecture with 124 billion total parameters, activating 5.1 billion per token. Ant said this structure is intended to combine the breadth of a large model with lower running costs and easier deployment.

It also reported competitive results across benchmarks including FinFIRST, FinSearchComp Verified, FinCRAFT, FinanceAgent v1.1/v2, APEX-Agents, SpreadsheetBench v1/v2 and τ3-Banking. These tests cover search, spreadsheet handling, reasoning and agent-based financial tasks.

Workflow focus

Ant said the system focuses on four areas of financial work: information retrieval, research reasoning, valuation modelling and report generation.

For information retrieval, the model is designed to prioritise official and authoritative sources so data can be checked from origin to output. In research reasoning, it is intended to combine information from multiple sources and formats into verifiable evidence chains.

Its valuation modelling function is aimed at handling complex Excel links and supporting automated updates while keeping files editable. The report-generation function is meant to combine facts, calculations and charts into professional research documents.

The release also includes open weights, allowing users to deploy the model privately and connect it to external tools such as search, Python, databases and spreadsheets. This lets firms adapt the model to internal financial processes rather than relying only on standard interfaces.

Benchmark release

Alongside the model, Ant is also open-sourcing FinFIRST, a benchmark for financial search agents developed with support from the investment banking team at China International Capital Corporation.

According to Ant, FinFIRST V1 contains 123 expert-authored tasks, 701 atomic criteria and 12,300 rubric points. The benchmark is designed to assess the full research process rather than only whether a final answer matches a reference output.

That approach reflects a broader issue in financial AI, where the process used to reach an answer can matter as much as the answer itself. Financial work often requires users to show where information came from, how calculations were made and whether each step can be reviewed later.

Ant positioned Ling-3.0-flash-Fin as a response to those demands, arguing that financial applications require systems that can handle shifting information, accounting standards and compliance requirements rather than operate only as general question-and-answer tools.

Wider model set

The launch is part of Ant's broader Ling 3.0 portfolio, which includes several models aimed at different use cases. Ling-3.0-flash is the base hybrid-reasoning Mixture-of-Experts model from which the finance version is derived.

Ant also offers Ling-3.0-tiny, a smaller model intended for local deployment without cloud dependence. Another variant, Ling-3.0-flash-VL, extends the model to image and video inputs and is aimed at tasks involving visual perception, document understanding and multimodal agents.

It has also introduced Ling-3.0-flash-Santé, a model tailored for healthcare and life sciences. This reflects a strategy of developing specialised models for sectors where domain-specific requirements shape how artificial intelligence systems are evaluated and deployed.

The open-sourcing of Ling-3.0-flash-Fin adds to a growing trend among large technology groups and model developers to release domain-focused systems with accompanying benchmarks. In finance, that trend has increasingly focused on tools that support research, analysis and document workflows while preserving traceability and control over data handling.

By pairing the model with a benchmark built around expert-authored financial tasks, Ant is signalling that performance in finance should be measured across the full chain of work. As it put it, FinFIRST evaluates "the full research process rather than relying solely on final-answer matching".