On November 3, Meituan officially launched LongCat-2.5, a new-generation native multimodal large model designed for long-horizon agent tasks. As a major iteration based on LongCat-2.0, the model has received a comprehensive upgrade in native visual understanding and agentic coding capabilities, and has competed head-to-head with top overseas large models in multiple authoritative tests, topping several metrics.

Core Architecture and Parameter Scale: 1.6 Trillion Total Parameters Empower End-to-End Engineering Capabilities

In terms of underlying architecture, LongCat-2.5 adopts a massive parameter system, with a total parameter count of up to 1.6 trillion, while activating about 48 billion parameters during each token inference, balancing strong inference compute with operational efficiency.

image.png

To give the model robust capabilities for handling complex tasks, the Meituan team introduced richer long-horizon trajectory data into long-sequence training and, through large-scale long-horizon reinforcement learning training, significantly improved the model's end-to-end engineering capabilities in tasks such as complex codebase development and complete project construction. At the same time, the model innovatively represents visual signals as multi-level discrete tokens, enabling the backbone model to seamlessly process visual and textual inputs in a unified token form. This deep integration of "native visual understanding" and "agentic coding" not only gives the model outstanding performance on Visual Coding tasks, but also provides solid technical support for complex interactive tasks such as creative design and 3D.

On the compute foundation supporting underlying training, the system has been deeply optimized for domestic compute environments, focusing on breaking through the technical bottlenecks of long-context training and W4A4 inference. Through refined training-inference numerical consistency tuning and long-context rollout optimization, it fully supports the efficient operation of large-scale long-horizon reinforcement learning.

Authoritative Benchmark Evaluation: Standout Performance in Tasks Such as Web Automation and 3D Generation

To comprehensively evaluate the model's overall performance, this evaluation covered three major categories—code, general tasks, and visual agents—including a total of 9 different benchmarks. In head-to-head competition with multiple popular top overseas large models, including Claude-Opus-5, GPT-5.6-Sol, GPT-5.5, and Claude-Opus-4.8, LongCat-2.5 delivered impressive results on multiple sub-leaderboards:

In the AutomationBench test, LongCat-2.5 scored 51.8 points, successfully surpassing Claude-Opus-5 (50.3 points) in the same category.

In the BrowseComp benchmark, the model ranked first with a high score of 91.9 points.

In the 3DCodeBench (text-to-3D project) test, the model scored 55.4 points, achieving a slight lead over similar competitors.

Overall, the evaluation results show that LongCat-2.5 demonstrates strong dominance in agent tasks such as web browsing, automated operations, and 3D generation; while in more traditional task categories such as code writing and terminal commands, although a small gap remains compared with leading overseas large models, it has firmly established itself in the global top tier.

With the official debut of LongCat-2.5, Meituan not only showcased its deep technical accumulation in foundational large models and embodied systems, but also further proved that domestic agent large models are fully capable of competing head-on with the world's most advanced products. At present, the official online experience and API access channels have been officially opened, making it convenient for developers and industry users to conduct in-depth trials.