The competitive landscape of large AI models in China has recently seen subtle changes. According to the latest information, Meituan has issued an internal notice to strictly restrict access to the "Doubao" large model. This move indicates that Meituan is increasingly shifting toward its internally developed solutions for AI infrastructure.
According to internal management requirements, all business teams must immediately conduct self-inspections on their current use of the Doubao large model. For businesses that still need to retain external model calls, they must submit detailed explanations of necessity and migration plans, and can only proceed after passing a separate approval process. In fact, this is not the first time Meituan has tightened control over external large models. Earlier this year, in April, Meituan had already stopped recommending the use of Alibaba Cloud's Qwen (Tongyi Qianwen) model for internal business use. Related usage applications now require approval from senior company leadership (X3 level), highlighting Meituan's emphasis on the controllability of its large model supply chain.
Meituan's confidence in "excluding" external models is closely tied to its gradually matured in-house large model ecosystem. As a core technology foundation, the LongCat large model, which was launched in 2023, has been deeply integrated into the "Xiaotuan" AI assistant and various merchant management tools. Recently, Meituan also officially released the LongCat-2.0 model with trillions of parameters, and simultaneously announced it would open-source the model.
From a technical perspective, LongCat-2.0 not only achieves full-process training on domestic computing clusters, but also has native support for ultra-long context of up to 1 million tokens. By covering multiple dimensions of data including Chinese, multilingual, and code processing, the model has become a core driving force for Meituan to enhance the intelligence level of its business. In addition, Meituan also introduced the LongCat-AudioDiT audio generation model and the multimodal model LongCat-Next, building a comprehensive AI capability matrix covering vision, speech, and text.
Industry observers point out that Meituan's series of actions, from restricting external models to fully promoting its in-house LongCat system, reflect deep considerations by major internet companies about the sovereignty of core AI technologies. With the iteration of in-house model capabilities and the construction of an open-source ecosystem, Meituan is trying to enhance the flexibility and security of its business through the establishment of a "technical closed loop" in the intense AI race. This also signals that the large model market in China is accelerating into a new phase where major players are focusing on strengthening their internal capabilities.
