The rapid growth of online programming learning platforms has created a strong demand for personalized exercise recommendation systems to enhance learners’ self-study efficiency. In response, sequential recommendation models have gained attention for their ability to leverage the temporal order of user interactions. This study conducts a comprehensive empirical evaluation of several sequential models using a real-world dataset from the CodePTIT programming learning platform. Among them, SASRecF achieves the best performance on the test set, leading in 8 out of the 9 evaluation metrics. The result also highlights that, by concatenating items and their features as input, the model effectively captures both sequential patterns and item-level information, leading to superior recommendation accuracy. These findings further demonstrate the potential of sequential modeling in enhancing recommendation quality for online programming education systems.