Ant Group releases open-weight financial model and open benchmark for financial search

Ant Group posted two open components on X intended for practical financial work: a model called Ling-3.0-flash-Fin and a benchmark called FinFIRST. The stated goal is to make financial AI more accessible and verifiable. The model uses an MoE (mixture-of-experts) architecture with 124 billion total parameters, of which 5.1 billion are active per forward pass, and a 256 thousand-token context window.

Because the weights are open, teams can run the model privately, connect it to search engines, Python environments, databases and spreadsheets, and adapt it to specific financial workflows. The model is available for download on Hugging Face, enabling immediate experimentation without dependence on a closed API or the provider's cloud.

FinFIRST is an open benchmark for financial-search agents, developed by Ant Group with professional guidance from the investment-banking team at China International Capital Corporation Limited (CICC). The first version contains 123 expert-written tasks, 701 atomic criteria and 12,300 detailed evaluation points, a volume designed to enable granular measurement of answer quality in a financial setting.

The company invites developers to join a dedicated Discord community for updates, deployment discussions, feedback and collaboration. Notably, the announcement uses the term "open-sourcing" but explicitly notes "open weights," an important technical distinction: the weights are open for use and adaptation, but the training code, data and full license do not necessarily meet the definition of fully open source.