Meituan opens 1.6-trillion-parameter LongCat AI model on Chinese chips

Meituan has released LongCat-2.0 and said the model was trained and deployed entirely on a 50,000-chip cluster built around Chinese-made processors, a claim that puts hardware sovereignty at the center of China’s latest AI push.

The Chinese delivery and local-services giant said the system carries 1.6 trillion total parameters, with about 48 billion activated per token and a dynamic activated range of roughly 33 billion to 56 billion. Meituan’s technical materials say the project began exploring domestic compute in 2023, giving the company a two-year runway before this release. That timing matters because the model is being presented not just as a bigger system, but as proof that frontier-scale training can be done without Nvidia hardware.

LongCat-2.0 is also built for long-horizon work. Meituan said it supports a native 1 million-token context window, which makes it suited to long documents, code understanding, generation and execution, and agentic coding tasks. The company released the model under the MIT License and posted it on GitHub and Hugging Face, a move that should make it easier for enterprises and developers to test and integrate it.

The launch lands in a crowded Chinese AI market where DeepSeek and ByteDance’s Doubao are already pressing for attention, while domestic chipmakers such as Huawei and Enflame are moving to cover gaps left by Western suppliers. Meituan’s own claim is the sharper one: that LongCat-2.0 is the first trillion-parameter model of its kind to complete full training and inference on a domestic compute cluster. That framing turns the release into a test case for China’s response to U.S. export controls, where access to top-end AI chips has become a strategic constraint rather than a routine procurement problem.

Meituan’s earlier AI systems were already tied to commerce. The company has used prior assistants to recommend restaurants and hotels and to help users order food or book rooms, and earlier reporting said a pilot customer-service agent improved operational efficiency by 20%. LongCat-2.0 extends that path toward what Meituan and other Chinese platforms increasingly call agentic commerce, where AI does not just search but completes transactions. Official and near-official materials also said the model preview ranked among the top three globally by call volume on OpenRouter before the public release, underscoring demand even before the weights were broadly available.

The release shows how quickly China’s AI race is shifting from model size alone to the infrastructure behind it. Meituan’s message is that the next contest is not only about who builds the most capable system, but who can build it on home-grown silicon.