Meituan releases LongCat-2.5, a 1.6 trillion parameter multimodal model for agent tasks

Meituan has officially open-sourced LongCat-2.5, a native multimodal model designed for long-horizon agent tasks. The release followed an initial announcement on November 2nd, with the open-sourcing made public on November 9th. The model is available under the MIT license, permitting commercial deployment and fine-tuning without additional licensing fees. LongCat-2.5 boasts a total parameter count of 1.6 trillion, with approximately 48 billion parameters activated per token, and supports a massive one-million token context window. To optimize for domestic computing chips with constraints on memory and bandwidth, the development team implemented deep collaborative optimizations across the model architecture, chip adaptation, and deployment strategies. Key technical advancements include transforming visual signals into multi-level discrete tokens that are processed uniformly by the main backbone, rather than relying on external visual modules. Furthermore, the model utilizes a W4A4 efficient inference scheme, quantizing both expert weights and activations to 4 bits for computation via native INT4 matrix multiplication units. The model demonstrates strong performance in Visual Coding, supporting tasks like creative design and 3D generation.