Meituan Open-Sources 1.6T-Parameter LongCat-2.0, Trained on Chinese ASICs

Meituan's AI team open-sourced LongCat-2.0 on June 30 under an MIT license, a Mixture-of-Experts model with 1.6 trillion total parameters and a native 1-million-token context window. Notably, the entire pipeline—from training to large-scale inference—was completed on a superpod of roughly 50,000 cards built from domestic AI-specific ASICs, with no reliance on NVIDIA GPUs. (overview)