In response to the challenges of predicting bearing limit of composite-bolted joints under the coupling effects of multi-source assembly factors， as well as the low efficiency of traditional numerical simulations，this paper proposes a rapid prediction method for the bearing limit based on the Vision Transformer （VIT） framework. By integrating the geometric deformation parameters and physical performance parameters during the assembly process，a nonlinear mapping model between the multi-source assembly parameter space and bearing limit is established. First，the geometric and performance parameters of the composite-bolted joint assembly process are analyzed and modeled. Then，an innovative CBJ-VIT deep learning network based on the VIT architecture is developed，utilizing a multi-head self-attention mechanism to achieve the feature fusion of multimodal assembly data. Finally，case studies are conducted with aerospace thin-walled composite-bolted assemblies for validation. The experimental results indicate that the predicted results from the CBJ-VIT model align closely with the finite element analysis results，reducing the time required for a single prediction from the traditional numerical simulation time of 12 h to just 8.1 s. In both qualitative and quantitative evaluations，the model shows an 85.02% improvement in the prediction accuracy compared with traditional non-image data processing methods and a 76.24% accuracy increase compared with non-VIT architecture models.