Real-world long-term time series are challenging to model due to their complex temporal patterns. While popular RNN-based, Transformer-based, or linear modeling approaches achieve excellent results in long-term time series forecasting, they often fail to capture both global and local views of the time series and ignore its intrinsic properties. This paper proposes CMNet, which combines the stochastic downsampling strategy (SDS) and the hybrid fusion of convolution and MLP (CM). Specifically, the SDS module addresses the intrinsic limitation of representation ability in one-dimensional time series by effectively downsampling them into a more compact format. Immediately, we use a hybrid method of convolution and MLP to extract information from the local and global perspectives of the sequence and then focus on potential relationships between the multiple variables. Experimentally, CMNet outperforms the state-of-the-art methods on five out of seven widely used benchmark datasets, what’s more, CMNet is less than them in terms of the number of parameters.