Meituan has unveiled LongCat-2.5, its next-generation native multimodal large model designed for long-horizon agent tasks. The model features 1.6 trillion total parameters, with approximately 48 billion activated per token. Across nine benchmarks spanning three major categories—coding, general tasks, and visual agents—LongCat-2.5 was evaluated against several leading overseas models and topped multiple leaderboards, including AutomationBench, BrowseComp, LongCatClawBench, and 3DCodeBench, demonstrating clear advantages in web browsing, automated operation, and 3D generation agent tasks, while placing in the top tier for coding and terminal instruction tasks. Building on LongCat-2.0, LongCat-2.5 enhances native visual understanding and agent programming capabilities through techniques such as the introduction of more long-horizon trajectory data and large-scale long-horizon reinforcement learning. These gains are underpinned by optimized domestic computing power training and inference infrastructure, demonstrating that domestic agent large models are now capable of competing on equal footing with top overseas products.