Meituan open-sourced the LongCat-2.5 model on November 9, with 1.6 trillion total parameters and a one million token context
LongCat-2.5 is released under the MIT license with commercial inference deployment and retraining allowed, represents visual signals as multi level discrete tokens processed by the same backbone as text, and uses W4A4 quantization for inference
On November 9, Meituan announced that the LongCat-2.5 model is officially open source. The model was released on November 2 and targets long-horizon agent tasks. It is published under the MIT license, allowing commercial inference deployment and retraining without additional authorization, with clearly improved end-to-end delivery on complex codebase development, full project builds, visual coding and professional office scenarios.
The model has 1.6 trillion total parameters, activates about 48 billion parameters per token and supports a one million token context. For domestic compute chips constrained in memory capacity and bandwidth, the team optimized the model architecture, chip adaptation and deployment strategy together. The visual capability gains come from reworking the visual representation layer, where the model represents visual signals as multi level discrete tokens processed by the same backbone as text rather than through an add on vision module. Visual tokens can be precomputed and cached offline, reducing online computation and making distributed training loads more even.
On the inference side, LongCat-2.5 uses a W4A4 efficient scheme that quantizes both expert weights and activations to 4 bits and computes through native INT4 matrix multiplication units. Model weights are already on the model hosting platform, with the technical blog and the API platform open at the same time.