Beijing-based Meituan said on June 30 it had released LongCat-2.0 as open source. The 1.6 trillion-parameter AI model was trained and deployed entirely on domestic hardware, it said.

Meituan said the model has a 1 million-token context window, putting it alongside DeepSeek-V4-Pro, which also lists 1M context and 1.6T parameters.

The company said LongCat-2.0 is part of its in-house foundation model program and is aimed at coding. Meituan said the broader LongCat family includes text, image, audio, and video models, as well as agents.

Meituan did not identify the chip vendor. It said training and inference ran on AI application-specific integrated circuit superpods. In a WeChat post, the company also said it used Huawei Collective Communication Library (HCCL) to improve training stability.

The launch adds to signs that Chinese AI companies are training advanced models on domestic hardware as US export controls limit access to leading AI chips.

🔗 Source: South China Morning Post

Recent Meituan developments