Benchtop nuclear magnetic resonance (NMR) devices enable rapid on-site detection of new psychoactive substances (NPS) at customs or mobile checkpoints, addressing the urgent need for real-time screening in combating illicit drug trafficking. Benchtop NMR systems typically exhibit low signal-to-noise ratios, posing challenges for accurate substance identification, particularly in complex mixture scenarios. Traditional machine learning models, despite their application in spectral analysis, struggle with these low signal-to-noise conditions and limited data sets, resulting in suboptimal performance for benchtop NMR-based NPS detection. We propose NMR4NPScreen, a deep learning model designed for NPS nontargeted screening using benchtop NMR data, capable of classifying nine distinct NPS categories with high accuracy. Our model adopts a channel attention-enhanced architecture combined with chemically informed preprocessing and contrastive pretraining that aligns NMR spectra with SMILES representations. This design substantially strengthens spectral feature extraction under low signal-to-noise conditions and enables chemically consistent embeddings, thereby overcoming the intrinsic limitations of benchtop NMR data. The model achieves an accuracy of 94.8%, surpassing traditional machine learning approaches, and demonstrates high robustness in detecting mixtures. This work paves the way for deploying advanced neural network models in NMR applications, enhancing real-time NPS detection capability.