LongCat-2.0: A 1.6T-Parameter MoE Model Built for Agentic Coding, Now on OpenRouter
Key Info
LongCat-2.0 is a 1.6T-parameter mixture-of-experts model with roughly 48B active parameters and 1M-token context support, built specifically for agentic coding. It is the full model behind Owl Alpha and is now available on OpenRouter.
Highlights
- Efficient long-context design: LongCat Sparse Attention (LSA) scales effectively to 1M-token contexts.
- Zero wasted compute: Dynamic expert activation ranges from 33B to 56B per token, avoiding unnecessary computation.
- Task-routed experts: MOPD uses three specialized expert groups — Agent, Reasoning, and Interaction — gated per task.
- Strong benchmark showing: Scores 70.8 on Terminal-Bench 2.1, with SWE-bench results also reported.