Meituan Releases LongCat-2.0, a 1.6T-Parameter MoE Model with Native 1M Context
Meituan has released LongCat-2.0, a large-scale Mixture-of-Experts (MoE) language model with 1.6 trillion total parameters, activating about 48 billion per token. This model is designed for agentic coding, encompassing code understanding, generation, and execution within agent workflows. It uniquely supports a native 1-million-token context window, achieved using LongCat Sparse Attention, and was trained entirely on domestic AI ASIC superpods, avoiding NVIDIA hardware for both training and serving.
LongCat-2.0 pushes the boundaries of context window size for efficient agentic coding and demonstrates the viability of large-scale AI model development and deployment on non-NVIDIA hardware, increasing competition and reducing dependency in the AI chip market.
Learn one new AI thing every day.
Daily Deck sends you seven plain-English cards like this every morning. Free.
Start free