Reportedly: Meituan LongCat releases LongCat-2.5 native multimodal AI model: 1.6T parameters, targeting long-horizon agent tasks
Model: LongCat-2.0 · availability, license and releases
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Meituan LongCat released LongCat-2.5, a native multimodal model with 1.6 trillion total parameters and about 48 billion activated per token, targeting long-horizon agent tasks. It builds on LongCat-2.0 with enhanced native visual understanding and agentic coding. The release was reported by IT Home on November 2, 2026.
China context
- Original name
- 美团龙猫
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- If weights are released, developers can fine-tune LongCat-2.5 for long-horizon agent tasks; otherwise, they must rely on Meituan's API or wait for open-source alternatives.
- For investors
- Meituan's investment in large-scale long-horizon RL and domestic compute optimization signals a strategic push into agentic AI, which could affect competitive dynamics in enterprise automation.
Translated from Chinese. Quotes and facts link to the original sources.
Meituan LongCat released LongCat-2.5, a native multimodal model with 1.6 trillion total parameters and about 48 billion activated per token, targeting long-horizon agent tasks. The model builds on LongCat-2.0 with enhanced native visual understanding and agentic coding. It incorporates more long-horizon trajectory data in long-sequence training and large-scale long-horizon reinforcement learning to improve end-to-end engineering capabilities for complex codebase development and full project construction. The training pipeline includes data synthesis, environment construction, and cross-harness mixed training. LongCat-2.5 represents visual signals as multi-level discrete tokens, allowing the backbone model to process visual and text inputs in a unified token form. This combination of native visual understanding and agentic coding leads to strong performance on Visual Coding and supports creative design and 3D tasks. The upgrades are supported by optimized long-context training on domestic compute and W4A4 inference, with numerical consistency between training and inference and long-context rollout optimization for efficient large-scale long-horizon RL.
LongCat-2.5 uses a 1.6T parameter architecture with about 48B activated per token, indicating a sparse mixture-of-experts design. It processes visual signals as multi-level discrete tokens, unifying visual and text modalities. Training includes long-horizon trajectory data and large-scale long-horizon reinforcement learning, with infrastructure optimizations for domestic compute, W4A4 inference, and long-context rollout.
Developers outside China may face a new open-weights competitor for long-horizon agent tasks if Meituan releases the model weights; otherwise, access is limited to Meituan's ecosystem. The model's focus on agentic coding and visual understanding could pressure existing coding agents and multimodal models. Verification needed: whether weights or API access are available outside China.
LongCat-2.5 could reduce development costs for complex software projects by automating codebase-level tasks, but its business value depends on availability and performance in real-world settings. Meituan may use it internally to improve its own services, such as food delivery logistics or customer support.
Watch for Meituan to release model weights or API access, and for independent benchmarks on Visual Coding and long-horizon agent tasks. The next signal is whether LongCat-2.5 appears on public leaderboards or is adopted by third-party developers.