Meituan releases new model with 1.6 trillion parameters, general agent test exceeds Opus-5, domestic core is specially optimized

Zhidongxi Author | Yang Jingli Editor | Li Shuiqing Zhidongzhi reported on November 2 that today, the Meituan LongCat team released LongCat-2.5, a native multi-modal model for long-range agent tasks. The total number of parameters of this model is 1.6 trillion, and each Token activates approximately 48 billion parameters. Based on LongCat-2.0, it focuses on strengthening native visual understanding and intelligent agent programming capabilities. In several benchmark tests such as the automation task benchmark AutomationBench and the general agent task benchmark LongCatClawBench announced by Meituan, LongCat-2.5 is ahead of Claude-Opus-5, GPT-5.6-Sol and other models; on the terminal operation benchmark Terminal Bench v4.0, LongCat-2.5 scored 38.5, More than GPT-5.6-Sol, but still behind Claude-Opus-5. LongCat-2.5 introduces more long-range trajectory data into long-sequence training, and improves the model's end-to-end engineering capabilities in tasks such as complex code base development and complete project construction through large-scale long-range reinforcement learning. At the same time, LongCat-2.5 converts visual signals into multi-level discrete Tokens, allowing visual and text input to be processed uniformly by the same backbone model, further supporting visual programming (Visual Coding), creative design and 3D content generation. Faced with the limitation of less than 80GB of video memory in a single card of domestic accelerators, LongCat-2.5 rearranges FSDP, EP, and CP parallelism around super nodes, superimposes hybrid vocabulary parallelism, MoonEP dynamic expert replication, and dual-shard FSDP to move visual coding out of the training calculation graph. LongCat said that after the above optimization, the single-card throughput of the FSDP configuration in long sequence SFT tasks increased by 86% compared with the DDP baseline. After the model was released, smart things also immediately focused on visual programming and office tasks.