The company paired the CatPaw platform launch with a major Xiaotuan upgrade that adds task execution for consumers, underscoring a broader AI push across both users and merchants.
Meituan denied market speculation on July 27 that Pei Peng, head of foundational models for the LongCat team, was preparing to leave the company, saying the claim was untrue. The rebuttal came as Meituan broadened its artificial intelligence push with the official launch of CatPaw, a full-scenario AI Agent platform, and a comprehensive upgrade to its local services AI assistant, Xiaotuan. As of press time, Meituan-W was trading at HK$90.35, up 4.21%, with turnover at HK$3.76 billion, or about $480.0 million. The company’s AI strategy now spans both consumers and merchants. Xiaotuan has moved beyond search, question answering and decision support to add agentic execution capabilities, allowing users to complete tasks such as placing orders, hailing rides and booking reservations through natural-language requests. Meituan said Xiaotuan’s information services cover more than 2,800 cities in China, with more than 700 million merchant information verifications and over 1.3 billion user reviews cataloged, giving it a large data base for local services matching and execution. On the merchant side, CatPaw is positioned as an out-of-the-box AI workstation and a platform for enterprise-grade agent development and hosting. Market reports cited in the source said it has already been deployed internally across 90,000 employees and hosts more than 30,000 agents, with validation completed in multiple real-world business scenarios including dining, beauty and veterinary clinics. Within user authorization and preset operating rules, the platform can aggregate sales, voucher redemption, reviews and inventory data, identify abnormalities and generate analysis and recommendations, while leaving critical decisions such as pricing changes and marketing campaigns to human operators. Meituan’s AI push has drawn attention since the release of LongCat-2.0 a month earlier. The company described it as the industry’s first trillion-parameter large model to complete training and inference entirely on domestic computing power. The model uses a MoE architecture with 1.6 trillion total parameters and activates about 48 billion parameters per token. It also supports a 1M ultra-long context window and is adapted to Claude Code, OpenClaw and Hermes for coding tasks. Pei Peng joined Meituan in 2023 after focusing for years on information retrieval and natural language processing. Public records show he first served as pre-training lead for the foundational model team in the computing and intelligent platforms division before becoming head of foundational models for the LongCat team. In that role, he has led research and development for Meituan’s large language models, multimodal models and AI agent models, including the LongCat series and the R&D and deployment of the trillion-parameter LongCat-2.0. Guotai Haitong Securities said recent growth could also be supported by three incremental drivers: instant retail, the overseas business KeeTa and AI-powered automated fulfillment. The report said instant retail benefits from Meituan’s existing food delivery network, while KeeTa is testing its profit model in Hong Kong and overseas markets and expanding into the Middle East and Brazil. It added that self-developed large models, autonomous delivery and embodied robots are helping shift Meituan from a pure transaction platform toward a service-industry technology operator.