China – Ant Group has open-sourced Ling-3.0-flash-Fin, an AI model purpose-built for financial workflows, according to an announcement made at the 2026 Inclusion·Conference on the Bund in Shanghai on 9 September.
The release marks the company’s latest step in bringing AI into practical, real-world scenarios, with a specific focus on delivering utility and efficiency for complex financial tasks.
Financial work depends on trustworthy sources, consistent definitions, accurate calculations and auditable outputs, and Ant Group said that AI in this sector must go beyond simple question-and-answer interactions.
Instead, the company said the technology needs to navigate dynamic information, complex accounting standards, and stringent compliance requirements.
To that end, Ling-3.0-flash-Fin was co-developed with leading financial institutions and industry experts, with professional expertise embedded directly into the model’s task definitions, data systems, and evaluation frameworks.
Ant Group said the model was guided by three core principles: being professional, efficient and open.
On the technical side, Ling-3.0-flash-Fin is built on a Mixture-of-Experts (MoE) architecture, with 124 billion total parameters, of which only 5.1 billion are activated per token.
According to Ant Group, this design allows the model to deliver the comprehensive knowledge capacity of a large model while maintaining the low inference costs and high deployment efficiency typically associated with a more compact system.
The company said this architectural efficiency translates into strong practical performance, with the model posting competitive results across a range of benchmarks.
These include FinFIRST, FinSearchComp Verified, FinCRAFT, FinanceAgent v1.1/v2, APEX-Agents, SpreadsheetBench v1/v2, and τ³-Banking.
Beyond raw benchmark performance, Ling-3.0-flash-Fin is designed around four core capabilities tailored specifically for investment research and financial analysis.
The model combines information retrieval—prioritizing authoritative sources for consistency and traceability—with research reasoning that synthesizes multi-source data into clear, verifiable evidence chains.
It also handles valuation modelling, understanding complex Excel financial linkages while keeping files fully editable, and report generation, integrating facts, calculations, and charts into professional research outputs.
Ant Group said Ling-3.0-flash-Fin forms part of a broader Ling 3.0 series, reflecting the view that different industries and scenarios require specialised solutions.
Among these is Ling-3.0-flash, a hybrid-reasoning MoE model built for production-scale agents, which also features 124 billion total parameters with 5.1 billion active per token.
According to Ant Group, this model achieves performance comparable to flagship 1 trillion-parameter models on most benchmarks while significantly reducing compute requirements.
Another member of the series is Ling-3.0-tiny, a native hybrid reasoning model with 7.9 billion total parameters and 1.3 billion active per token, designed for resource-sensitive deployment.
Ant Group said Ling-3.0-tiny runs entirely locally without cloud dependency, making it suited to personal knowledge management and offline tasks.
The series also includes Ling-3.0-flash-VL, a vision-language model built on Ling-3.0-flash that adds image and video inputs alongside a 1 million token context window.
Ant Group said this model is designed for visual perception, STEM reasoning, document intelligence, multimodal agent tasks and medical report interpretation.
Rounding out the portfolio is Ling-3.0-flash-Santé, an MoE model enhanced for healthcare and life sciences applications.
According to Ant Group, this model achieves top-tier performance among flash-size models on key medical benchmarks including MedXpertQA-Text and DiagnosisArena-MCQ, delivering flagship-level performance for clinical reasoning and evidence-based retrieval.
Ling-3.0-flash-Fin is now available through OpenRouter and Vercel, with its open weights accessible via Hugging Face and ModelScope.
Ant Group said users can deploy the model privately, connect it to search tools, Python, databases and spreadsheets, and adapt it to their own financial workflows.
Alongside the model, Ant Group is also open-sourcing FinFIRST, an expert-built benchmark designed specifically for financial search agents.
FinFIRST was developed with professional support from the investment banking team at China International Capital Corporation Limited (CICC). The first version, FinFIRST V1, includes 123 expert-authored tasks, 701 atomic criteria and 12,300 rubric points.
Rather than relying solely on final-answer matching, Ant Group said FinFIRST evaluates the full research process, a design intended to ensure strict data consistency and end-to-end traceability in complex financial scenarios.
