Meituan Releases LongCat-2.5: A 1.6T Parameter Native Multimodal Model for Long-Horizon Agents
Meituan's LongCat team has released LongCat-2.5, a native multimodal large model with 1.6 trillion total parameters. By leveraging large-scale long-horizon reinforcement learning and multi-level discrete visual tokens, the model significantly enhances end-to-end engineering capabilities in complex codebase development, Visual Coding, and long-horizon agent tasks. Online access and APIs are currently available, with open-source weights coming soon.
Full Report
IT Home reported on November 2 that Meituan LongCat has released LongCat-2.5, a native multimodal model designed for long-horizon agentic tasks. The model features a total parameter count of 1.6 trillion, with approximately 48 billion parameters activated per token. Building upon LongCat-2.0, it further enhances native visual understanding and agentic coding capabilities.
According to the introduction, LongCat incorporated more long-horizon trajectory data into its long-sequence training and leveraged large-scale long-horizon reinforcement learning to improve end-to-end engineering capabilities in complex tasks such as developing large codebases and building complete projects. The official team continues to expand this training pipeline around data synthesis, environment construction, and cross-Harness mixed training. Additionally, LongCat transforms failure trajectories from real-world agents into reproducible and verifiable training tasks, continuously supplementing training data to address specific model weaknesses.
Meanwhile, LongCat-2.5 represents visual signals as multi-level discrete tokens, enabling the backbone model to process both visual and textual inputs in a unified token format. The combination of native visual understanding and agentic coding allows the model to excel in Visual Coding, while also supporting tasks such as creative design and 3D generation.
The officials stated that these upgrades are supported by continuously evolving infrastructure: optimizing long-context training and W4A4 inference on domestic computing power, and ensuring efficient execution of large-scale long-horizon RL through training-inference numerical consistency and long-context rollout optimizations.
IT Home provides the following evaluation results:
- Online Experience: https://longcat.chat
- API Access: https://longcat.chat/platform/product
- GitHub, HuggingFace: Coming Soon
Advertisement Disclaimer: External links contained in this article (including but not limited to hyperlinks, QR codes, passcodes, etc.) are provided to convey additional information and save selection time. Results are for reference only. All IT Home articles containing external links include this disclaimer.