Meituan’s LongCat-2.5 runs 49B of its 1.6T parameters

Meituan has put LongCat-2.5, a multimodal model that activates 49B of its 1.6T parameters per token, on its APIs, the company said in a November 2 post on the Meituan LongCat blog. If you’re considering it for coding or long agent runs, keep in mind that the comparison with GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8 and Claude-Opus-5 comes from Meituan’s own announcement.

LongCat-2.5 is a native multimodal model with a 1M-token context window. Meituan built it for long-horizon agentic tasks and coding, and says it improves on visual understanding and coding. Its total parameter count matches LongCat-2.0, the mixture-of-experts model Meituan released before it, which runs about 48B active parameters per token.

In the LongCat-2.0 repository, Meituan scored its previous model at 59.5 on SWE-bench Pro. That only edged out GPT-5.5 at 58.6, and it trailed Claude Opus 4.8 at 69.2. Those rival scores carry an asterisk, which in the repo means they were not measured in-house under Meituan’s unified harness. LongCat-2.0’s own score was measured there.