LongCat-2.5 opens multimodal agent weights
Meituan's LongCat-2.5 model card describes a native multimodal mixture-of-experts model for visual understanding, code engineering and long-horizon agent tasks. It reports 1.6 trillion total parameters and roughly 48 billion activated per token under an MIT license. The maker reports 88.4 on Terminal Bench v2.1 and labels its in-house scores as indicative. Open weights enable controlled evaluation, while hardware, latency, tool permissions and real-task reliability still require separate testing. A benchmark score does not guarantee unattended project delivery.
SourceMeituan LongCat