Meituan unveils LongCat-2.0, a 1.6T-parameter AI model trained on domestic Chinese chips

Meituan has unveiled LongCat-2.0, a self-declared 1.6 trillion-parameter AI model designed for agentic coding that was fully trained and deployed on a 50,000-card cluster of domestic Chinese processors, marking a significant claim for China's AI self-reliance amid export restrictions. The model features a 1-million-token context window and activates approximately 48 billion parameters per token, having been trained on over 35 trillion tokens while overcoming large-scale communication and memory challenges without recovery failures. Although Meituan positions LongCat-2.0 as a proof point for operational independence from Western chips, the company has not publicly identified the specific chip supplier or model, leaving a key detail of this technological breakthrough unconfirmed. This development matters to the global AI industry and Chinese tech sector as it challenges assumptions about the viability of domestic hardware for frontier-scale models, even though independent benchmarks suggest the model's performance may still lag behind top Western counterparts.