Time series forecasting has been extensively researched in extensive domains such as traffic, finance, and industry. Due to the evolving and intricate nature of time series dataset, forecasting tasks are highly challenging. To address this issue, a plethora of models based on CNNs, RNNs, or attention have been developed. However, related works have pointed out that while these models perform well in forecasting tasks, they often overlook the impact of inputs at different scales on model training. In this paper, we analyze the multi-scale architecture and the linear mixer model, proposing MSMixer, a novel approach that combines multi-scale modules and multilayer perceptron mixer modules. With the introduced multi-scale module, MSMixer is capable of handling inputs at different scales for the same length, thereby capturing more information. Each mixing block embedded with the MLP mixer module isolates external influences, effectively capturing dependencies between patches and channels. Through comprehensive experimentation utilizing 7 real-world time series datasets, MSMixer demonstrates outstanding performance compared to various state-of-the-art models (8 baselines), showcasing its efficiency and generalization capability.