On November 2, Meituan (03690) Longcat released LongCat-2.5, a native multimodal model built for long-horizon agent tasks.
Then, arriving today (the 9th), Meituan (03690) announced that the model has been officially open-sourced.
The model uses the MIT license and supports commercial inference deployment and secondary training without any additional authorization, and it is a native multimodal model designed for long-horizon agent tasks.
It delivers clear improvements in end-to-end delivery capability across complex codebase development, full project construction, Visual Coding, and professional office scenarios.
The model has 1.6 trillion total parameters, activates roughly 48 billion parameters per token, and supports a one-million-token context.
For domestic computing power chips constrained by video memory and bandwidth, the team also carried out deep collaborative optimization in model architecture, chip adaptation, and deployment strategy.
According to the introduction, LongCat-2.5 performs outstandingly on Visual Coding and supports tasks such as creative design and 3D.
These capability gains come from a reworking of the visual representation layer: the model represents visual signals as multi-level discrete tokens, which, like text, are processed uniformly by the backbone model rather than through an external visual module.
Specifically, the model applies residual vector quantization to native-resolution visual features, supplementing details level by level; visual tokens can be precomputed offline and cached, reducing online computation and making distributed training loads more balanced; the visual codebook and the backbone are trained jointly, with vision and text sharing the same set of backbone parameters, collaboratively completing visual understanding and code generation within the same semantic space.
In addition, LongCat-2.5 adopts a W4A4 efficient inference scheme, quantizing both expert weights and activations to 4 bit and executing computation through native INT4 matrix multiplication units.
Comments