Infrared small target detection (IRSTD) is a critical yet challenging task due to low target-background contrast, minimal target texture, and high noise levels. While data-driven methods have significantly advanced performance, persistent challenges remain. Traditional pooling schemes in multi-scale networks often sacrifice crucial details necessary for detecting small targets, and existing attention mechanisms struggle to balance performance with computational efficiency. To address these challenges, we propose the Hybrid Feature Mining Network (HFMNet), a lightweight and efficient model designed to enhance feature representation for IRSTD. First, to reduce detail loss caused by conventional pooling—which can amplify noise (e.g., max pooling) or suppress contrast (e.g., average pooling)—we introduce the Cross-Pooling Shallow Enhancement Module (CSEM). By integrating diverse pooling strategies across parallel shallow streams, CSEM preserves fine details, improves localization, and prevents target loss in deeper layers. Second, for efficient local-global context modeling, we propose the Hybrid Feature Mining Module (HFMM). This module combines local and global attention by decomposing large 2D convolutions into coupled 1D convolutions and incorporating dilated convolutions, broadening the receptive field while reducing complexity. Coupled with the Vision State Space Module (VSSM) for global context modeling, HFMM effectively integrates local and global information to enhance detection performance. The effectiveness of our HFMNet has been evaluated on four benchmark datasets. In comparison to other SOTA methods, the Pd reaches 92.52% on the IRSTD-1K dataset and 99.08% on the NUAA-SIRST dataset, with only 7.09M parameters, making it a promising lightweight solution for IRSTD applications. The code will be released at https://github.com/Fortuneteller6/HFMNet.