Foundation models, pivotal to the next generation of artificial intelligence, are beginning to reshape the financial sector by lowering development barriers and enabling more intelligent systems.
At the World Artificial Intelligence Conference (WAIC) on July 28, Ant Digital Technologies unveiled Agentar-Fin-R1, a large model designed for financial reasoning.
Agentar-Fin-R1 is built specifically for financial applications. It achieved top scores across three key industry benchmarks, reportedly outperforming both general-purpose open-source models and specialized financial ones such as DeepSeek. According to the company, the model excels in domain expertise, reasoning ability, and compliance safeguards.
To support varied deployment needs, Ant also released non-reasoning versions of its financial models with 14 billion and 72 billion parameters, along with a mixture-of-experts (MoE) model based on the Bailing foundation model.
Despite recent advancements, foundation models continue to face domain-specific challenges in finance, including hallucination, output inconsistency, and limited explainability. These limitations underscore the relevance of Ant’s new financial reasoning model.
“Rather than sticking with generic models, we’ve evolved toward specialized solutions,” said Wang Wei, CTO of Ant Digital Technologies, in an interview with 36Kr. “We’re drawing deeply on years of experience working closely with clients, scenarios, and problem-solving approaches, which is enabling a true vertical shift.”
Ant has developed a detailed task taxonomy for financial services, reportedly covering six major categories and 66 subcategories across banking, securities, insurance, funds, and trusts. The model is trained on hundreds of billions of financial tokens and integrates a chain-of-thought (CoT) framework guided by domain experts—enhancements that improve its performance on complex reasoning tasks.
For training, Ant employs a weighted algorithm that boosts learning efficiency and task-specific performance while reducing the data volume and computing power needed for downstream fine-tuning. This approach helps lower deployment costs.
Beyond model development, Ant is building an end-to-end technology stack that spans foundation models, AI platforms, and finance-specific business applications.
If the large model serves as the “brain,” the financial AI agent acts as the “body,” translating cognitive capacity into real-world execution. Agentar-Fin-R1 is key to accelerating this transformation.
To date, Ant said it has worked with financial institutions to launch more than 100 agentic AI solutions across banking, securities, insurance, and broader financial services. According to the company, these systems have improved frontline staff efficiency by over 80%.
Technical innovation alone will not enable foundation models to deliver business outcomes. Deep knowledge of financial workflows is just as crucial, and Ant benefits from having years of operational experience on both fronts.
“Our firm sits at the intersection of AI and finance,” said Zhang Peng, head of AI at Ant Digital Technologies. “We’re uniquely positioned. Ant Group has tested these agents in its own operations, while Ant Digital Technologies works closely with external financial institutions. That perspective gives us a rich, cross-cutting view of what works.”
Ant is not alone in this effort. At WAIC 2025, financial AI agents were prominently showcased by technology firms of all sizes. The industry is moving beyond pilot projects and into core functions like credit approval, marking a shift from experimentation to scale.
“We’re entering an era where AI agents are in full bloom,” Wang said. “But it’s a marathon, not a sprint. We plan to keep running deep into the verticals, especially in finance, where Ant’s strengths can really shine.”
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Wang Fangyu for 36Kr.