Meituan Releases 1.6-Trillion-Parameter AI Agent Model LongCat-2.5 Trained on Chinese Hardware
The Chinese company Meituan has made the model weights of its LongCat-2.5 publicly available. The AI agent-focused system is said to have been trained entirely on Chinese hardware.
The Chinese delivery and services group Meituan has made the model weights of its new AI system LongCat-2.5 publicly accessible. The model has 1.6 trillion parameters and is designed for use in AI agents, meaning software that can carry out tasks independently. What stands out in particular is that, according to the announcement, training took place entirely on Chinese-developed hardware.
The announcement does not mention Nvidia graphics processors. As an example of practical deployment, a configuration with 194 chips from Huawei's Ascend series is cited. The total number of chips used for training and the overall amount of computing effort are not disclosed in the information available so far. It therefore remains unclear to what extent the project actually relied on domestic hardware.
Meituan reports a score of 88.4 points for LongCat-2.5 on Terminal-Bench, a test of agent capabilities. On this benchmark, the model thus trails Claude Opus 5 by a narrow margin. An independent verification of this result was not yet available at the time of publication, so the performance figure currently rests on the company's own statements.
The release comes at a time when Chinese technology companies are increasingly presenting capable models while addressing their dependence on leading Western hardware. For observers, therefore, the system's performance is not the only point of interest; the question of how reliable the claims about the training infrastructure are also matters, as long as they cannot be independently verified.
Source: the-decoder.com
59
