LongCat-2.0
Meituan · 2026-06-30 · 1.8T parameters
LongCat-2.0 is Meituan's open-weight large-scale MoE language model, marketed as a 1.6T model with about 48B parameters activated per token; its learned checkpoint tensors contain 1,775,560,457,088 parameters after excluding persistent routing-correction buffers. It was pretrained on more than 35 trillion tokens and introduces LongCat Sparse Attention, N-gram embedding expansion, 1M-context training data, and post-training focused on coding, agentic workflows, search, and optional reasoning.