BEIJING — Chinese food delivery giant Meituan on Tuesday unveiled LongCat-2.0, a next-generation large language model that it said is the world’s first trillion-parameter artificial intelligence system trained and operated entirely on a 50,000-chip computing cluster powered by Chinese-made processors.
The company also announced it would open-source the model, underscoring China’s accelerating push to develop advanced AI technologies using domestic hardware as U.S. export restrictions continue to limit access to high-end American chips.
Often compared with DoorDash, Meituan entered China’s competitive AI race later than rivals such as DeepSeek and ByteDance’s Doubao. Its LongCat development team was established in 2023 and introduced its first model only late last year.
Meituan did not disclose how LongCat-2.0 would be deployed across its core businesses. Earlier versions of the model have powered AI assistants within the company’s platform that recommend restaurants and hotels while helping users complete tasks such as ordering meals and booking accommodation, part of the growing trend toward so-called agentic commerce.
The company also suggested the model could support applications beyond its delivery ecosystem. In a statement published on LongCat’s official WeChat account, Meituan highlighted the AI system’s ability to create gaming websites and write novels, signaling potential expansion into broader commercial AI services.
The launch comes as Chinese technology companies increasingly seek to reduce reliance on foreign semiconductor suppliers following U.S. export controls introduced in 2022.
Major AI developers, including DeepSeek, Alibaba and ByteDance, have accelerated efforts to train models using domestic computing infrastructure, while Chinese chipmakers such as Huawei and Enflame have expanded their presence through supply agreements with AI firms.
According to Meituan, LongCat-2.0 was trained from scratch on a cluster containing 50,000 Chinese-made AI chips and is capable of processing inputs of up to 1 million tokens, allowing it to analyze exceptionally long documents.
The company said the model was specifically optimized for agentic coding, with an architecture designed to improve performance on complex real-world software development tasks.
Meituan added that a preview version of LongCat-2.0 had already become one of the three most-used models on OpenRouter, a widely used global marketplace for AI models.
The company also claimed LongCat-2.0 matched or outperformed several leading proprietary AI systems, including Google’s Gemini, OpenAI’s GPT-5.5 and Anthropic’s Claude Opus, on selected coding and autonomous agent benchmarks.
“LongCat-2.0 has demonstrated that we now have the capability to train large-scale models on domestic computing clusters,” Meituan said in its statement, without identifying the manufacturer of the processors used in the system.
