Meituan has released LongCat-2.5, an advanced AI agent model boasting 1.6 trillion parameters, developed and trained entirely on Chinese hardware without the use of Nvidia GPUs, according to The Decoder.
The model scored 88.4 points on the Terminal Bench benchmark, placing it just behind Claude Opus 5 in performance. Notably, the deployment sample runs on 194 Huawei Ascend chips, reflecting a significant push for domestic AI infrastructure.
This development signals growing momentum in China’s AI sector, which is closely monitored by Japanese investors and tech firms seeking to understand shifts in the competitive landscape for AI hardware and software.
