To address the difficulty and high time-complexity of retinal vessel segmentation, an asymmetric model called compact mixed network(CMNet) is proposed, which is capable of achieving trade-off between speed and accuracy. Firstly, considering the ability of deformable convolution to extract complex and variable vascular structure, and that large kernel in mixed depthwise convolution can further improve segmentation quality while increasing the receptive field, we propose a lightweight mixed bottleneck module. Then, an adaptive feature layer fusion is proposed to further improve the spatial mapping capability of the model. Finally, the vessel segmentation performance is analyzed quantitatively and qualitatively. The AUC metrics are 0.9840, 0.9879 and 0.9853 for DRIVE, CHASE_DB1 and HRF benchmark datasets, respectively, indicating that the proposed algorithm is able to obtain highly accurate segmentation results. Furthermore, with an input resolution of 512times512, the model achieves a frame rate of 33 FPS on a single V100 GPU, which further indicates its suitability for rapid clinical deployment.