A giant agent model trained on Chinese hardware
Meituan has published the weights of LongCat-2.5, a model designed for AI agents. Its most notable aspect is that no NVIDIA GPUs were used in training. The company says the sample deployment runs on 194 Huawei Ascend chips.
1.6 trillion parameters, built for agents
With 1.6 trillion parameters, LongCat-2.5 is one of the largest open-weight models to date. It is specifically focused on agentic tasks, meaning software that can make decisions and carry out multi-step actions on its own.
88.4 points on Terminal Bench
The model scores 88.4 on the Terminal Bench evaluation, placing it just behind Claude Opus 5. The result shows that a Chinese open-weight model is approaching the top tier in agent capabilities.
The NVIDIA dependence debate
The absence of any mention of NVIDIA GPUs is notable at a time when chip supply and export restrictions are pushing companies toward alternative hardware. An open-weight model running on Huawei Ascend chips is a concrete example of Chinese AI firms strengthening their own hardware ecosystem.
