Meituan's LongCat-2.5: a 1.6T agent model tested inside Claude Code
A food delivery company just posted a 1.6 trillion parameter agent model. Meituan's LongCat-2.5 is a mixture-of-experts model with about 48 billion active parameters, natively multimodal, and pitched squarely at long-horizon agent work. The blog went up dated November 1.
On LongCat's own runs it scores 88.4 on Terminal Bench v2.1, 64.6 on DeepSWE v1.1, 51.8 on AutomationBench and 91.9 on BrowseComp, the highest BrowseComp figure in its comparison table. It still trails on the hardest coding evals: 38.5 on Terminal Bench v4.0 against 53.9 reported for Opus 5, and DeepSeek-V4.1-Flash is listed above it on both Terminal Bench v2.1 and DeepSWE. Competitor numbers come from public reports, not reruns.
Two details are more interesting than the scores. Coding evals ran inside the Claude Code harness, version 2.1.280, and the model was trained with "mixed training across harnesses". Meituan also says it converts real-world agent failure trajectories into reproducible, verifiable training tasks. That is the harness-is-the-model idea turned into a training recipe: if your users run agents in Claude Code, train in Claude Code. It also runs W4A4 inference on domestic Chinese accelerators.
The catch: GitHub and Hugging Face are both marked "coming soon", and no license is stated. Until the weights and license show up, this is an API model with a strong agentic table, not an open release. When they do, read the license before you build on it.
Blog: https://longcat.ai/blog/longcat-2.5/
← Back to all articles
On LongCat's own runs it scores 88.4 on Terminal Bench v2.1, 64.6 on DeepSWE v1.1, 51.8 on AutomationBench and 91.9 on BrowseComp, the highest BrowseComp figure in its comparison table. It still trails on the hardest coding evals: 38.5 on Terminal Bench v4.0 against 53.9 reported for Opus 5, and DeepSeek-V4.1-Flash is listed above it on both Terminal Bench v2.1 and DeepSWE. Competitor numbers come from public reports, not reruns.
Two details are more interesting than the scores. Coding evals ran inside the Claude Code harness, version 2.1.280, and the model was trained with "mixed training across harnesses". Meituan also says it converts real-world agent failure trajectories into reproducible, verifiable training tasks. That is the harness-is-the-model idea turned into a training recipe: if your users run agents in Claude Code, train in Claude Code. It also runs W4A4 inference on domestic Chinese accelerators.
The catch: GitHub and Hugging Face are both marked "coming soon", and no license is stated. Until the weights and license show up, this is an API model with a strong agentic table, not an open release. When they do, read the license before you build on it.
Blog: https://longcat.ai/blog/longcat-2.5/
Comments