Aiming to address the issues of excessive parameters, poor real-time performance, and insufficient lightweight design in traditional models for soybean field weed recognition, an improved lightweight weed detection model named YOLO 11n - ADS was proposed based on YOLO 11n. By replacing the original C3K2 module with a switchable atrous convolution (SAConv) module and integrating a global context mechanism with multi-dilation-rate feature fusion, the enhanced model strengthened multi-scale feature extraction capability and improved weed detection accuracy in complex environments. Meanwhile, the backbone network was optimized by using the adaptive downsampling module (ADown), which fused the average pooling and maximum pooling strategies to reduce the feature loss caused by target occlusion and light interference. The YOLO 11n - ADS model achieved detection accuracy, mAP50, mAP50-95, and recall rates of 90.6%, 92.0%, 86.0%, and 92.6%, respectively, representing improvements of 1.1 percentage points, 2.6 percentage points, 5.0 percentage points, and 3.6 percentage points over the baseline YOLO 11n model. Additionally, the parameter count was reduced from 2.6×10^6 to 2.2×10^6, while the floating point operations (FLOPs) was decreased from 6.3×10^9 to 4.6×10^9. Deployed on the Jetson Nano edge computing platform by using the TensorRT framework, the optimized model achieved efficient real-time inference with minimal resource consumption. The memory usage was only 1.6 GB, and the detection speed reached 35 f/s. In field dynamic validation tests, the end-to-end detection average frame rate reached 28 f/s, with detection accuracy, mAP50, mAP50-95, and recall rates of 89.5%, 91.0%, 84.0%, and 91.4%, respectively. Compared with YOLO 11n, the proposed model demonstrated stronger robustness and real-time performance under challenging scenarios such as uneven lighting and target occlusion. The research result can provide an efficient and lightweight solution for real-time weed detection in complex farmland environments, which was suitable for deployment on edge computing devices and contributed to the intelligent management of precision agriculture.