LongCat-2.0: The Stealth AI Model That Was Quietly Topping OpenRouter All Along
Meituan unveiled LongCat-2.0, an open-license 1.6T-parameter mixture-of-experts model that had been running anonymously on OpenRouter as “Owl Alpha.” It activates about 48B parameters per token and was trained and deployed end-to-end on more than 50,000 domestic Chinese accelerators, making it a major domestic ASIC-based training milestone. Pretraining used over 35T tokens and reportedly completed without rollbacks or major stability failures. The model targets agentic coding and long-context use, with 1M-context support, sparse attention, a richer N-gram embedding scheme, and a router that combines specialist systems for tools, reasoning, and chat. Benchmark results include 59.5 on SWE-bench Pro and 73.2 on FORTE, competitive with leading models in some tasks. Its main advantage is cost: standard API pricing is $0.75/$2.95 per million input/output tokens, temporarily discounted to $0.30/$1.20, with free cached reads and cheap token packs. It is available via API and agent harnesses, but self-hosting weights are not yet released.
