At Bund Conference, Ant Group Lays Out AI Stack From Ling 3.0 Models to Robot Brains

At the conference, the AI took over concrete tasks rather than sitting inside a chat window. Visitors verified their identity by tapping their phones, an action that also woke the exhibition's AI assistant. The most conspicuous exhibit was a robot moving through pharmacy aisles only 80 centimeters wide, taking orders, identifying and sorting medicines and handing them over, with no renovation of the pharmacy or redesign of its shelves required. The robot runs on Lingbo, Ant's general-purpose brain for embodied systems.

Ant's two main consumer products are A'Fu and A'Bao. A'Fu, aimed at health, has 150 million users and handles about 20 million health consultations a day, linking registration, triage, report interpretation and post-diagnosis follow-up; when a village doctor sees a patient, the conversation can be turned into a medical record automatically, with diagnostic suggestions shown alongside and a referral recommended when a case exceeds the system's scope. A'Bao covers dining, mobility, travel, culture and government services, with more than 10,000 services adapted for AI. A user can ask it to buy a cup of coffee every morning at 10; the agent remembers the arrangement, finds the service, creates the order and completes payment.

For companies, Ant Digital Technologies has upgraded its Agentar platform into what it calls a super factory for commercial agents. Agentar comes with roughly 200 job-level digital expert templates and several hundred industry skills that businesses can assemble into agents for customer service, marketing, risk control and research. Zhao Wenbiao, chief executive of Ant Digital Technologies, said at the conference that the unit's overall business had grown by close to 50 percent a year over the past two years and that its AI-to-business revenue was growing at triple-digit rates. Ant also showed Xiaoyu, an operations agent that answers questions such as which store is selling poorly or which goods need restocking, and MYbank's Bailing 2.0, which lets owners call on credit, bill and tax capabilities in everyday language.

The models beneath these products are the Ling 3.0 series. Ling-3.0-flash has 124 billion total parameters but activates 5.1 billion per token, while the lightweight Ling-3.0-tiny has 7.9 billion total parameters and activates 1.3 billion. Ant offers multimodal and finance-oriented variants, Ling-3.0-flash-VL and Ling-3.0-flash-Fin. According to Leiphone, Ling-3.0-flash was trained in more than 10,000 interactive environments, with an emphasis on tool calling and the execution of complex tasks, a design meant to keep inference costs low enough for the models to be used in real operations.

Lingbo carries the same logic into the physical world. Ant did not build its own robot first; it supplies a brain that can drive machines of different shapes, demonstrated in pharmacies, logistics warehouses and on industrial production lines.

Ant's argument for differentiation rests less on the breadth of that portfolio than on depth in finance and healthcare, two sectors where data sits in separate systems and where permissions, compliance and accountability constrain what a model may do. Payment and security are presented as part of the same landing capability. AI Pay, APASS and KYA are intended to establish which agent is acting, whom it represents and under what authority, and whether an action can be reversed or traced. "Capability decides what AI can do; trust decides what society dares to hand to AI," said Chen Liang, Ant Group's security chief technology officer.

Han Xinyi said at the conference that the ChatGPT moment for agent commerce has arrived. Asked by the moderator whether that judgment came from Ant's backend data, he replied: "We actually haven't seen such signs from our backend data." Users want agents to handle more tasks, but merchants have not finished adapting their services, and without visible demand they are reluctant to invest ahead of it. Models can iterate quickly; industry does not move at the same speed.