Aerial object detection aims to precisely identify and spatially localize ground targets from aerial images, an essential task for a wide range of applications. However, atmospheric haze significantly degrades traditional algorithms, limiting their practical utility. Prior methods employing image dehazing as preprocessing face a mismatch between low-level restoration and high-level detection objectives. This paper introduces Hazy-Mamba, an end-to-end detection framework for UAVs in hazy scenes. It enhances feature extraction and detection using the Haze-Adaptive Feature Enhancement (HAFE) and Multi-scale Attention Gating (MAG) modules. HAFE extracts dehazed features through low-pass filtering, multi-scale convolution, and contrast enhancement, while MAG addresses blurry features and small targets via a bottleneck structure and multi-scale extraction. Additionally, Hazy-Mamba incorporates an Adaptive Haze-aware Distillation Loss, based on knowledge distillation, which combines haze-aware weights and KL divergence loss to dynamically adjust the learning focus. This approach achieves a balance between accuracy and complexity. Experiments demonstrate that Hazy-Mamba attains improved mAP on the HazyDet dataset with a reduced parameter count, rendering it suitable for UAV deployment.