1.6T total / ~48B active per token (MoE)
1M tokens
MIT
Official open-weight LongCat flagship release positioned for coding and agentic workflows, with public Hugging Face weights, a published chat template, and deployment guidance for both GPU and NPU environments. The official model card describes LongCat-2.0 as a 1.6 trillion-parameter Mixture-of-Experts model with about 48 billion activated parameters per token, trained on more than 35 trillion tokens and hundreds of billions of 1M-context tokens on AI ASIC superpods rather than Nvidia GPU infrastructure. Release materials emphasize LongCat Sparse Attention, which introduces streaming-aware indexing, cross-layer indexing, and hierarchical indexing to reduce long-context indexing overhead, plus a 135B-parameter N-gram Embedding module inherited from LongCat-Flash-Lite to improve parameter utilization outside the MoE path. Meituan positions the model as deeply compatible with Claude Code, OpenClaw, Hermes, OpenCode, and related developer harnesses, targeting repository-level edits, automated task execution, search, and long-horizon coding work. The official benchmark tables highlight in-house or reported results across Terminal-Bench 2.1, SWE-bench Pro, SWE-bench Multilingual, FORTE, BrowseComp, RWSearch, IFEval, Writing Bench, IMO-AnswerBench, and GPQA-diamond, presenting LongCat-2.0 as competitive with leading proprietary agentic and coding models while remaining locally deployable with published SGLang support.