Run LongCat-2.0 in Python: Meituan's Owl Alpha

For weeks, one of the highest-volume models on OpenRouter had no name. Developers knew it only as Owl Alpha — an unlabeled endpoint that kept winning blind coding comparisons. On June 30, 2026, Meituan pulled off the mask: Owl Alpha was LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts model trained end to end on a 50,000-card cluster of domestic Chinese accelerators, with no Nvidia hardware in the loop.

The reveal was a rare event in LLM marketing: a model that earned its ranking anonymously, on output quality alone, before anyone knew who built it or where it came from. It posts 59.5 on SWE-bench Pro (ahead of GPT-5.5 and Claude Opus 4.6 on that benchmark), 70.8 on Terminal-Bench 2.1, and it serves a native 1M-token context window through a technique Meituan calls LongCat Sparse Attention.