Deep learning-based methods have dominated in object detection in synthetic aperture radar (SAR) images. Despite significant advancements, existing approaches mainly focus on architectural enhancements of the network, leaving the unique challenges posed by the strong speckle noise inherent in SAR imagery not fully tackled. In this article, we introduce the multilevel denoising detection transformer (MD-DETR) to mitigate the impact of speckle noise for SAR object detection. We build MD-DETR in three steps, leveraging the advanced real-time detection transformer (RT-DETR) framework. First, we address the noise data input by designing a straightforward image-level denoising technique, which concatenates the original image with its denoised counterparts to create a new image for model input. Subsequently, for feature-level denoising, a coarse-mask guidance feature learning module is proposed to enhance the global features of the decoder. In addition, for query-level denoising, we introduce an auxiliary head with the one-to-many matching strategy. This enables flexible query selection by dynamically adjusting the number of queries to suit various scenarios. To validate the superiority of the proposed MD-DETR, extensive experiments are conducted on two benchmark datasets, that is, the SAR ship detection dataset (SSDD) and the recently published COCO-level large-scale multiclass SAR object detection dataset (SARDet-100K). Experimental results on both datasets outperform previous advanced detectors, achieving a new state-of-the-art (SOTA) with 98.9 AP50 and 88.5 mAP50 on SSDD and SARDet-100K, respectively.