Meituan has introduced the LongCat-2.5 model with 1.6 trillion parameters

Chinese company Meituan has released the weights for the LongCat-2.5 model, which features 1.6 trillion parameters. The architecture is designed specifically for autonomous AI agents and was trained entirely on domestic Chinese hardware without the use of Nvidia accelerators.

In the Terminal Bench benchmark, the new model scored 88.4 points, placing it in the rankings just behind the Claude Opus 5 model. A cluster of 194 Huawei Ascend chips was required for the system's test deployment.

The release demonstrates the capabilities of independent large-scale training of heavy language models on alternative hardware. The open weights allow developers to use the solution for local and specialized agent tasks.