Meituan releases 1.6 trillion parameter agent model LongCat-2.5, with no Nvidia GPUs used throughout training
Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU
Meituan's LongCat team released the weights of the agent model LongCat-2.5, with 1.6 trillion total parameters and about 48 billion activated per token, and it is said to have used Chinese chips throughout training. The model supports text, image and video input, converts images to tokens at native resolution for processing, and has an API context of up to 1 million tokens and output of up to 128,000 tokens. Training used reinforcement learning on real GitHub PR tasks, with shortcuts such as internet access and Git history blocked in the sandbox, and a language model checks whether solutions take shortcuts. According to evaluations published by the LongCat team, it scored 88.4 on Terminal Bench v2.1, close to Claude Opus 5's 89.1; DeepSWE rose from 9.7 in the previous generation to 64.6; BrowseComp scored 91.9, the highest in the table, but was measured by Meituan in its own environment.