This study evaluated whether inertial measurement units (IMUs) and deep-learning models could automatically identify improper squat technique based on the standards set by the National Academy of Sports Medicine (NASM). Twenty uninjured athletes (half male, half female) were asked to perform six repetitions each of five variations of the squat movement, one of which with proper form, and the rest with various form errors involving the knees and lower back. Five body-mounted IMUs were placed on the athletes and used to collect motion data and train various architectures including convolutional neural networks (CNN), gated recurrent units (GRU), a CNN–GRU hybrid, and the TabNet-enhanced versions of the three models. It was found that the GRU and GRU-TabNet models showed the lowest performance, and the CNN–GRU model performed best, reaching 99.7% classification accuracy with five sensors and 98.8% accuracy using only a single right-thigh sensor. Despite these high accuracies, the sample size was small and had little diversity, so no confidence intervals or external validation was provided. This limited the certainty of the findings.