Magnetic resonance imaging (MRI) is a vital medical imaging modality, but its development has been limited by prolonged scanning time. Deep learning (DL)-based methods, which build neural networks to reconstruct MR images from undersampled raw data, can reliably address this problem. Among these methods, model-driven DL methods incorporate different prior knowledge into deep networks, thereby narrowing the solution space and achieving better results. However, the complementarity among different prior knowledge has not been thoroughly explored. Most of the existing model-driven networks simply stack unrolled cascades to mimic iterative solution steps, which are inefficient and their performances are suboptimal. To optimize the conventional network structure, we propose a collaborative model-driven network. In the network, each unrolled cascade comprised three parts: model-driven subnetworks, attention modules, and correction modules. The attention modules can learn to enhance the areas of expertise for each subnetwork, and the correction modules can compensate for new errors caused by the attention modules. The optimized intermediate results are fed into the next cascade for better convergence. Experimental results on multiple sequences showed significant improvements in the final results without additional computational complexity. Moreover, the proposed model-driven network design strategy can be easily applied to other model-driven methods to improve their performances.