Pan Xin, former head of the AI platform at ByteDance's visual large model, has joined this week and is now in charge of multimodal AI innovation business, directly reporting to Meituan's Technology Committee. With this appointment, the Hong Kong-dollar-valued lifestyle service giant has completed the most critical piece of its "Delivery + AI" strategy for 2025.

Pan Xin's career spans top research systems in China and the U.S.: During his time at Google Brain, he led the dynamic graph module of TensorFlow, establishing the flexible training paradigm that has become the mainstream in global deep learning frameworks. After returning to China, he worked at Baidu, Tencent, and ByteDance, responsible for PaddlePaddle performance optimization, the construction of the WuLiang framework, and the visual large model platform. In November 2024, he joined Shangji Tech as an AI partner, driving the mass production of smart glasses. A source said that his technical management combines a framework-level perspective with product launch speed, which is exactly the "hard currency" Meituan urgently needs in its fierce delivery competition.

According to internal plans, Pan Xin's first task is to introduce the multimodal large model into the entire on-demand delivery process. The recently completed LongCat App can now synchronize rider trajectories, merchant food preparation, and traffic flow data, generating a "digital twin" for each order in the cloud, and predicting the optimal route in advance. Test data shows that the model can reduce the average fulfillment time by 2.8 minutes without increasing capacity, corresponding to 3-4 additional orders per rider per day.

Starting from the second quarter of 2025, Meituan has listed "basic model training" as its highest priority project. HR sources revealed that the company continues to expand its recruitment of AI training talent, targeting teams from Alibaba Tongyi, Seed, and Tencent Huan Yuan. Interview criteria focus on "real-world data loop" and minute-level decision-making ability. The already launched Nocode programming tool and the LongCat-Flash-Chat large language model are aimed at business managers and customer service scenarios, supporting natural language generation of scheduling scripts and automated responses to consumer inquiries.

At last week's strategic meeting, the management issued clear targets: the delivery business should improve human efficiency by 5% and reduce energy consumption by 3% using AI in the second half of the year, while ensuring that core operating cost growth remains below revenue growth. Industry analysts point out that as subsidy space shrinks, platform competition is shifting from capital to algorithm. "Whoever can use models to save time will seize the next opportunity." Pan Xin's joining provides Meituan the possibility to create a technological gap with competitors, but whether it can realize the model's benefits in the three experiences of riders, merchants, and consumers still needs continuous observation.