The model introduces LongCat Sparse Attention with streaming-aware, cross-layer, and hierarchical indexing to reduce the cost of ultra-long inputs. It also uses a 5-gram Embedding module, three-step multi-token prediction, and a 1M-token context training and serving stack. Post-training combines agent, reasoning, and interaction experts through a multi-expert architecture.
LongCat-2.0 is useful for repository-level coding, automated task execution, long-document analysis, and tool-enabled assistants. Developers can connect it through OpenAI-compatible or Anthropic-compatible API endpoints and integrate it with tools such as Claude Code, OpenClaw, Hermes, OpenCode, and Kilo Code; API access is billed by token usage.
