TMTPOST — Chinese internet giant Meituan on Tuesday officially released and open-sourced its next-generation LongCat-2.0 artificial intelligence model, featuring over one trillion parameters. The large-scale mixture-of-experts architecture—with a total capacity of 1.6 trillion parameters and an average of 48 billion active parameters per token—was trained from scratch on 50,000 domestically produced accelerator cards. The pre-training process consumed more than 30 trillion tokens spanning Chinese, English, multilingual text, and source code, enabling native support for a 1-million-token context window. Major technology platforms across the Chinese Mainland are increasingly turning to open-source frontier models as a way to challenge proprietary systems and accelerate ecosystem development. Meituan’s LongCat-2.0 release, optimized for native hardware, strengthens its position in long-context agentic workflows and software engineering tools within its consumer ecosystem. Demand for such massive, locally trained models is growing rapidly as enterprises seek greater control over data sovereignty and customization, potentially reshaping how AI infrastructure is built and monetized in the coming years.