LongCat-2.0

Meituan released LongCat-2.0, a large-scale Mixture-of-Experts language model with 1.6 trillion total parameters and roughly 48 billion activated per token, designed around agentic coding tasks such as code understanding, generation, and execution inside agent workflows. The model supports a native 1-million-token cont The architecture is built around four cost-reducing ideas, with the headline feature being LongCat Sparse Attention — which combines Streaming-aware Indexing, Cross-Layer Indexing, and Hierarchical Indexing to convert fragmented memory access into coalesced HBM reads, amortize indexing across adjacent layers, and apply

The official Meituan model card on Hugging Face introduces LongCat-2.0 as a 1.6-trillion-total-parameter MoE language model with 48 billion parameters activated per token, representing a substantial step up from previous LongCat models and accompanied by several architectural improvements. Both the full training run an To handle long-horizon tasks, the card introduces LongCat Sparse Attention (LSA), which makes three orthogonal improvements: Streaming-aware Indexing reshapes the token-selection budget into hardware-aligned reads for coalesced HBM access and high effective bandwidth; Cross-Layer Indexing amortizes a single indexing pa