HDMixer is a pure MLP-based model designed for multivariate time series (MTS) forecasting, aiming to enhance semantic information in patches and efficiently model hierarchical interactions. The model introduces a Length-Extendable Patcher (LEP) to adaptively extend patch lengths, preserving boundary information and reducing semantic incoherence. It also employs a Hierarchical Dependency Explorer (HDE) composed of MLPs to capture short-term, long-term, and cross-variable dependencies. The LEP dynamically adjusts patch lengths based on temporal characteristics, while the HDE uses stacked MLPs to model interactions across different dimensions. Extensive experiments on nine real-world datasets show that HDMixer achieves competitive performance, with top-1 results in 59 settings and top-2 in 13 settings. The model's efficiency is validated through performance, training speed, and memory usage comparisons. HDMixer outperforms state-of-the-art methods in terms of MSE and MAE, achieving significant reductions in error. The model's design ensures effective information preservation from long time series and efficient computation through shared parameters across dimensions. The LEP and HDE components are crucial for enhancing semantic information and capturing complex interactions in MTS forecasting.HDMixer is a pure MLP-based model designed for multivariate time series (MTS) forecasting, aiming to enhance semantic information in patches and efficiently model hierarchical interactions. The model introduces a Length-Extendable Patcher (LEP) to adaptively extend patch lengths, preserving boundary information and reducing semantic incoherence. It also employs a Hierarchical Dependency Explorer (HDE) composed of MLPs to capture short-term, long-term, and cross-variable dependencies. The LEP dynamically adjusts patch lengths based on temporal characteristics, while the HDE uses stacked MLPs to model interactions across different dimensions. Extensive experiments on nine real-world datasets show that HDMixer achieves competitive performance, with top-1 results in 59 settings and top-2 in 13 settings. The model's efficiency is validated through performance, training speed, and memory usage comparisons. HDMixer outperforms state-of-the-art methods in terms of MSE and MAE, achieving significant reductions in error. The model's design ensures effective information preservation from long time series and efficient computation through shared parameters across dimensions. The LEP and HDE components are crucial for enhancing semantic information and capturing complex interactions in MTS forecasting.