Meituan launched LongCat-2.5, a large model with total parameters of 1.6 trillion and about 48 billion active per token, supporting a 1-million-token context and native multimodality, AI-industry trackers reported on November 4. The model was trained entirely on domestic compute and targets long-horizon agents, visual programming and complex engineering tasks.
The release adds another major Chinese player to the trillion-parameter club and emphasizes practical agent capabilities rather than headline parameter counts. Meituan's move fits its position as a services platform where agent workflows can be applied to scheduling, operations and local-life commerce at scale.
Together with a batch of new model and chip announcements around the Wuzhen summit, LongCat-2.5 illustrates how China's AI industry is broadening from language models to domain agents built on homegrown hardware and infrastructure.
Source: Technology Media / Meituan
