Meituan LongCat releases LongCat-2.5 native multimodal model, 1.6T parameters targeting long-horizon agent tasks
Meituan LongCat releases the native multimodal model LongCat-2.5, with a total parameter count of 1.6 trillion and approximately 48 billion parameters activated per token, strengthening native visual understanding and agentic coding capabilities on the basis of LongCat-2.0. The model incorporates more long-horizon trajectory data into long-sequence training, and through large-scale long-horizon reinforcement learning enhances end-to-end engineering capabilities such as complex codebase development and complete project construction, while representing visual signals as multi-level discrete tokens to uniformly process visual and textual inputs. The official statement says it has optimized long-context training and W4A4 inference on domestic compute, and it will soon be available on GitHub and HuggingFace.
Why read this
AI score 68
The original text provides the 1.6T parameter and activation scale, and explains the specific approach of the long-horizon agent training pipeline, from which its engineering capability orientation can be judged.
Topics
Body · AI translation
Meituan Longcat Releases LongCat-2.5 Native Multimodal AI Model: 1.6T Parameters, Aimed at Long-Horizon Agent Tasks
2026-11-02 09:18 IT Home - Qincang (Intern)
Thanks to IT Home user GreatMOLA for the tip-off!
IT Home, November 2 news: Meituan Longcat today released LongCat-2.5, a native multimodal model aimed at long-horizon agent tasks. The model has a total parameter count of 1.6 trillion, with approximately 48 billion parameters activated per token, further strengthening native visual understanding and agentic coding capabilities on the basis of LongCat-2.0.
According to the introduction, Longcat added more long-horizon trajectory data to long-sequence training and, through large-scale long-horizon reinforcement learning, improved the model's end-to-end engineering capabilities in tasks such as complex codebase development and complete project construction. Around data synthesis, environment construction, and cross-Harness mixed training, the official team continues to expand this training pipeline. In addition, Longcat converts real Agent failure trajectories into reproducible, verifiable training tasks, continuously supplementing training data for the model's capability shortcomings.
At the same time, LongCat-2.5 represents visual signals as multi-level discrete tokens, enabling the backbone model to process visual and text inputs in a unified token form. The combination of native visual understanding and agentic coding capabilities gives the model outstanding performance in Visual Coding and supports tasks such as creative design and 3D.
The official team stated that these upgrades are supported by continuously evolving infrastructure: optimizing long-context training and W4A4 inference on domestic compute, and supporting the efficient operation of large-scale long-horizon RL through training-inference numerical consistency and long-context rollout optimization.
IT Home attaches the evaluation results:
Github, HuggingFace: coming soon
Advertising statement: External redirect links contained in the article (including but not limited to hyperlinks, QR codes, passcodes, and other forms) are used to convey more information and save selection time. The results are for reference only. All articles on IT Home that contain external links include this statement.