Chinese food delivery giant Meituan announced Tuesday that it has released its next-generation artificial intelligence model, LongCat-2.0, and plans to make it open-source. The company claims the system is the world’s first trillion-parameter AI model that was both trained and operates entirely on a cluster of 50,000 Chinese-made chips.
Meituan, which is frequently compared to the American app DoorDash, is a relative newcomer to China’s competitive and heavily funded AI industry. Its rivals include DeepSeek and ByteDance’s Doubao. The LongCat team was established in 2023 and only rolled out its first model late in 2024.
The company has not yet explained how LongCat-2.0 will be folded into its current services, though earlier versions of the technology have already been put to work powering in-app AI assistants that suggest restaurants and hotels and handle tasks like placing food orders and booking accommodations. This approach is part of a broader “agentic commerce” movement that rival Alibaba has also been pushing this year.
With consumer spending sluggish and profit margins tightening, Meituan appears to be looking for new ways to generate revenue. In a post on LongCat’s official WeChat account, the company highlighted the model’s capacity to build a gaming website and write a full novel.
The development of LongCat-2.0 using entirely Chinese chips reflects a broader push within China’s tech sector toward self-reliance. Companies including DeepSeek, Alibaba, and ByteDance have all been working to cut their dependence on American chips following export restrictions put in place by Washington starting in 2022. Chinese chipmakers have moved swiftly to fill that void, picking up business through supply agreements with AI developers.
According to the company, LongCat-2.0 was built from the ground up using those 50,000 domestic chips and is capable of processing inputs of up to one million tokens, enabling it to work with extremely long documents. The model is designed with a focus on agentic coding — meaning it is built to handle real-world programming tasks with greater efficiency and reliability.
Meituan also noted that a preview version of the model had already become one of the three most-used models on OpenRouter, a widely used global AI marketplace.
The company claims LongCat-2.0 matched or outperformed several prominent proprietary AI models — including Google’s Gemini, OpenAI’s GPT-5.5, and Anthropic’s Claude Opus — on certain coding and agent performance tests.
“LongCat-2.0 has demonstrated that we now have the capability to train large-scale models on domestic computing clusters,” the company stated, though it declined to identify which chipmaker supplied the processors.
