Meituan, a leading local service platform in China primarily known for food delivery, has achieved a significant milestone: it trained the artificial intelligence model LongCat-2.0, which possesses 1.6 trillion parameters. This model was entirely trained using domestic AI chips.
LongCat-2.0 is the world's first trillion-parameter model to undergo training and inference on chips manufactured in China. The scale of this model places it among the largest language models globally. However, an even more important aspect is what this demonstrates about the Chinese AI ecosystem and the unexpected advantages provided by operating a food delivery platform.
Meituan processes tens of millions of orders daily. Each order requires dynamic real-time routing for hundreds of couriers, vendors, and customers in complex urban environments. For instance, during peak lunch hours in Beijing within a three-kilometer radius, hundreds of couriers and thousands of orders are simultaneously active. The system must account for numerous variables, such as restaurant preparation times, traffic accidents, elevator failures, and address changes, requiring millisecond-level reoptimization.
This real logistics data, accumulated over more than a decade, constitutes a training resource that cannot be replicated by synthetic datasets. The architecture of LongCat-2.0 is optimized precisely for this reality: its 1 million token context window allows it to process entire codebases for agent coding tasks, and the trillion-parameter scale enables it to handle the combinatorial explosion of edge cases in real-time dispatching.
The deployment of 50,000 domestic AI chips marks a breakthrough for the Chinese semiconductor ecosystem. Before LongCat-2.0, the prevailing opinion was that domestic chips were not yet ready for training at the scale of advanced systems. Coordinating the operation of 50,000 chips, where the failure of a single component could lead to the collapse of the entire training process, required overcoming immense engineering challenges.
Meituan's motivation extends beyond simple delivery optimization. The base LongCat model is used to power internal AI agents responsible for code generation (EvoCUA), as well as for search and recommendation systems and self-evolving operations. By owning the entire stack—from silicon to the model itself—Meituan keeps confidential operational data within the company while simultaneously enhancing capabilities that could redefine the potential of its platform.
Furthermore, the existence of a trillion-parameter model signals a broader trend: Chinese internet giants like ByteDance, Alibaba, Tencent, and now Meituan are developing their own AI capabilities based on their proprietary data reserves. For Meituan, the delivery business provided the necessary data, and this data enabled the creation of a model that could potentially transform every aspect of the company's operations.
