Site-specific weed management relies on precise, real-time perception during early crop vegetative stages to optimize post-emergent herbicide usage and mitigate yield penalties. However, published object detection benchmarks rarely report species-level accuracy or detail computational performance under field-relevant hardware constraints. To address this, this study benchmarked 22 scale configurations across five leading detection architectures (YOLOv10, YOLO11, YOLO12, YOLO26, and RT-DETR) to identify eight economically significant weed species in commercial soybean fields. The dataset comprised 7,478 UAV-acquired RGB image tiles (containing 36,285 bounding box annotations) captured at 7 m above ground level. Models were evaluated using a rigorous five-fold stratified cross-validation protocol and an isolated independent test set. Averaged across all eight species, the detectors achieved 84.1% mAP@50 and 48.9% mAP@50-95. Crucially, performance across distinct architectural families varied by a mere 4.0 percentage points (p.p.), whereas accuracy across weed species fluctuated by 24.0 p.p., establishing target morphology and phenological complexity as the primary limiting factors. Variance decomposition attributed 97.3% of the predictive variability directly to the weed species, compared to only 1.7% for the detector choice. The YOLO12s model optimized the accuracy-latency Pareto frontier, delivering 49.9% mAP@50-95 with an inference latency of just 2 ms per image. Conversely, RT-DETR models exhibited slower convergence and inferior bounding-box localization for small, overlapping targets. Furthermore, confidence distribution analysis confirms that applying a 0.50 threshold effectively mitigates false positives with a minimal recall penalty, thereby establishing a highly robust and computationally efficient operational baseline for autonomous, localized smart-spraying systems.