Meituan Releases LongCat-2.5: A 1.6-Trillion Parameter Multimodal Model
Meituan has released LongCat-2.5, a massive 1.6-trillion parameter multimodal model, on Hugging Face. The model features native multimodality, advanced visual understanding, and strong capabilities for long-horizon agentic tasks. This release represents a significant development in the open-weights AI space, bringing massive-scale models to the community. It highlights the industry trend toward creating models optimized for complex engineering, multi-step planning, and autonomous AI agents. LongCat-2.5 utilizes a Mixture of Experts architecture with approximately 48 billion parameters activated per token. It is specifically designed to excel at agentic coding, tool use, and long-horizon task execution.
## BACKGROUND
Meituan is a major technology company known for developing advanced AI systems like the LongCat model series. Mixture of Experts (MoE) is a neural network architecture that activates only a subset of parameters for any given input, enabling massive model sizes while maintaining computational efficiency during inference.