Meituan opens LongCat-2.5 weights under the MIT license
China’s Meituan has released LongCat-2.5 model weights under the MIT license, allowing commercial inference and further training without separate authorization. The release follows the model’s earlier introduction and includes a company technical explanation of how visual and text inputs share a backbone. Its intended uses include extended coding tasks and professional office work.
Artificial Intelligence··Evening
The release opens commercial reuse
China’s Meituan, a technology company developing the LongCat model family, has released LongCat-2.5 weights under the MIT license. The terms permit commercial inference deployment and additional training without separate authorization. The release follows the model’s November 2 introduction: the new development concerns access to its weights. Independent coverage dated November 9 describes that opening, followed by Meituan’s November 10 technical announcement.[1], [2]
Visual and text inputs share a backbone
Meituan describes a multimodal architecture in which visual information and text are processed by the same backbone. Image features become discrete representations built in stages, with each stage adding detail. The visual representations can be computed in advance and cached. The company also says the visual codebook is trained together with the backbone, placing visual processing inside the model’s shared architecture.[1]
The model targets extended work tasks
The developer positions LongCat-2.5 for extended agent tasks, including work on complex code repositories, complete projects and professional office documents. Its technical description discusses quantization of expert weights and activations for inference, as well as coordination of computation and communication in distributed systems. The release makes weights available for reuse; claims about task improvements and deployment efficiency come from the developer’s own assessments.[1]
