蚂蚁数科的 Agentar 企业级全栈智能体平台，通过五大核心支撑构建起金融领域智能应用的可信底座， 既突破了金融场景的专业性、复杂性壁垒，又确保了技术应用的合规性与可靠性，最终推动金融新质生 产力实现跃迁。 支撑一：企业级全栈智能体平台，夯实技术底座 平台以 "1000 + 安全合规水位标准" 为基础，提供从底层架构到上层应用的全栈能力，支撑金融智能体 在复杂业务场景中稳定运行。其核心价值在于打通技术与业务的衔接，让智能体能够适配银行、保险、 证券等多类金融机构的差异化需求，为后续功能落地提供坚实的基础设施支撑。 支撑二：金融大模型，构建智能中枢 蚂蚁数科金融大模型是平台的 "大脑"，具备可靠、可控、可优化三大特性： 相比通用大模型，其在金融领域的语言理解、知识储备、逻辑推理和数学计算能力更突出，通过二次训 练(结合蚂蚁侧高质量金融数据与客户侧专有数据)可形成机构专有模型； 依托专业知识工程，构建了精标长 COT 数据、金融标签体系等高质量训练数据集，确保模型 "懂规 则、精业务"，减少 "幻觉"，保障数值计算准确与逻辑自洽。 支撑三：金融知识工程，破解专业壁垒 通过标准化加工金融知识资产，让智能体具备深度 ...

文｜光锥智能 当AI大模型已开始走向千行百业之时，备受看好的金融行业，却似乎陷入了落地瓶颈。 打开手机银行想查下贷款额度，对着屏幕说了半天，AI客服却只回复 "请点击首页贷款按钮"； 客户经理想用大模型生成一份客户资产配置方案，结果推荐的产品与客户风险等级完全不符； 风控团队测试的AI模型，在审批中小企业贷款时频频给出"幻觉答案"...... 这些看似荒诞的场景，却是当前AI落地金融行业时的真实困境。 当金融机构满怀期待地将AI请进门，却发现它既读不懂复杂的信贷政策，算不清理财产品的费率结构，更搞不懂不同银行的"行话体系"。 通用大模型的"聪明"，在严肃的金融场景里似乎失灵了，大模型与金融场景之间，也仿佛横亘着一道看不见的鸿沟。 "企业和产业需要的不是实验室的技术，而是真正能够解决真实问题的可信生产力。"蚂蚁数科CEO赵闻飙在2025WAIC智能体驱动产业变革论坛上如此说 道。 也正因此，面向AI大模型在具体行业中的落地，想要真正释放AI价值，关键就在于要从水平通用转向垂直专用，只有深度理解行业的大模型，才能懂行 话，做行活儿，创造真实的业务价值。 金融行业的AI落地，更是如此。 因此，构建专业的金融大模型是 ...

据悉，该模型专为金融业应用打造，在三大主流金融基准测试中均取得最高评分，超越Deepseek等开源 通用大模型及金融大模型，在金融专业性、推理能力以及安全合规能力上表现突出。 此外，蚂蚁数科还提供非推理版本的14B与72B参数金融大模型以及基于百灵基础大模型的MOE架构大 模型满足金融机构在多样化场景部署的需求。 作者 | 王方玉 编辑 | 苏建勋 AI大模型作为新一代人工智能的基石，为金融行业的智能化提供了强大的技术底座，降低了开发和应 用的门槛，堪称金融智能化的"新基建"。 7月28日，在世界人工智能大会（WAIC 2025）上，蚂蚁数科正式发布金融推理大模型Agentar-Fin-R1。 值得一提的是，蚂蚁数科不仅研发行业模型，还布局了从金融行业大模型、AI平台到上层应用的全栈 式解决方案。 打个比方，如果说大模型是"大脑"，那么智能体就是将这个"大脑"的认知能力转化为具体金融业务执行 力的"身体"和"行动者"。蚂蚁数科发布的金融推理大模型，有助于进一步加速金融智能体的落地应用。 据悉，目前蚂蚁数科已联合金融行业伙伴推出超百个金融智能体解决方案，覆盖银行、证券、保险、通 用金融等四大领域，金融机构可以" ...

Core Insights - The article emphasizes that Qifu Technology is becoming a key partner in the digital transformation of Chinese banks by integrating financial intelligence with business scenarios [1][7] - Qifu Technology's financial intelligence platform, Deepbank, and its various AI applications are designed to meet the real business needs of banks [2][7] Group 1: Financial Intelligence Development - Qifu Technology launched its self-developed financial intelligence platform Deepbank, which includes four core applications: AI Marketing Assistant, AI Approval Officer, AI Decision Assistant, and AI Compliance Assistant [2][3] - The financial intelligence system is built on a heterogeneous large model platform and a multi-agent collaborative framework, enabling deep understanding of financial semantics and user behavior [2][4] Group 2: Comprehensive Credit Solutions - The upgraded "Qifu Credit Super Intelligent Agent" includes five modules that cover the entire credit business process, enhancing decision-making capabilities [3][6] - The "End-to-End Credit Decision" module utilizes over 700 models and more than 1 billion historical decision data points to accurately assess user risk [3][4] Group 3: Technological Empowerment - Qifu Technology employs four technological engines: data flywheel, multi-modal fusion, self-evolution, and multi-agent collaboration to enhance the intelligence of its financial agents [4][5] - The AI Approval Officer can automatically extract key information from user applications and provide instant recommendations, significantly improving efficiency [5][6] Group 4: Practical Applications and Collaborations - Qifu Technology has demonstrated the effectiveness of its financial intelligence through partnerships with banks, leading to improved customer acquisition and reduced compliance risks [6][7] - A strategic cooperation agreement was signed with Guangdong Huaxing Bank to explore deep applications of AI in credit business, showcasing a collaborative model for digital transformation in banking [6][7]

Core Viewpoint - The importance of safety and trustworthiness in digital financial innovation is emphasized, highlighting the need for regulatory clarity and market confidence [1][5][6]. Group 1: Digital Financial Innovation Principles - The "high, medium, initial, small" principle is proposed to balance innovation and risk tolerance, allowing for some flexibility in risk management [1][8]. - Financial models must prioritize safety, reliability, and explainability, with a focus on advanced security technologies to prevent malicious attacks [2][4]. Group 2: Key Considerations for Financial Models - Financial models should avoid pitfalls such as model hallucination, discrimination, algorithmic resonance, AI deception, and the coldness of machine interactions [4]. - The need for legal clarity regarding the status and responsibilities of financial AI is highlighted, ensuring that financial institutions have clear decision-making accountability [4][6]. Group 3: Economic Efficiency and Collaboration - Industry-level financial models should be developed through extensive data pre-training and customization to reduce development costs and expand application ranges [5]. - Collaboration between strong tech companies and financial institutions is encouraged to lead the development of industry-level financial models and applications [5]. Group 4: Regulatory Innovation - The necessity for a robust regulatory framework for digital finance is stressed, including the establishment of clear business norms and a comprehensive regulatory system [6][7]. - The balance between innovation and regulation is crucial, with a call for a flexible approach that does not stifle innovation while ensuring market stability [7][8].