Meituan trained AI model LongCat-2.0 with 1.6 trillion parameters entirely on Chinese chips

Chinese company Meituan announced that its new AI model LongCat-2.0 with 1.6 trillion parameters became the first model of such scale trained entirely on domestic Chinese chips—a move the company directly calls a response to U.S. export restrictions.

What Is Known About the LongCat-2.0 Model

The most notable aspect of Meituan's new model is not its size, though it is indeed large, but the hardware on which it was trained. The company claims that LongCat-2.0 became the first model of such scale trained end-to-end on domestic Chinese chips, rather than on processors from Western manufacturers.

• Model name — LongCat-2.0, with 1.6 trillion parameters.
• The model, according to the company, is the first of this scale to undergo full training on domestic Chinese chips.
• Meituan directly links this move to U.S. export restrictions on advanced chips.

Why This Is a Response to U.S. Export Restrictions

Training a model with 1.6 trillion parameters entirely on domestic hardware is a technically complex task, typically requiring chips optimized for a specific architecture and compute volume. That Meituan managed to complete the entire process end-to-end without using Western chips, the company presents directly as a demonstration that U.S. export restrictions do not stop the development of large Chinese AI models. The framing of the announcement itself—not as routine technical news, but as a "pointed answer," that is, a direct and targeted response to export control policy—shows that Meituan views this news first and foremost in geopolitical rather than purely product context.

What Distinguishes LongCat-2.0 From Other Large Models

Most models of comparable scale worldwide continue to be trained using chips from Western manufacturers, which remain the industry standard for training large language models. Meituan's claim that LongCat-2.0 is the first model of such size trained entirely on domestic chips sets it apart on this basis—the company emphasizes not the architectural features of the model, but rather the infrastructure independence of the training process.

What This Means

Meituan's announcement is not simply news of another large language model, but a signal that Chinese technology companies are capable of scaling trillion-parameter AI models on their own chip base. If such claims are verified and repeated by other players, the significance of U.S. export restrictions as a tool for restraining Chinese AI will gradually diminish, and competition in developing the largest language models will finally cease to depend on a single chip ecosystem.