Mayer (2017, 2020) proposed three major design features of computer-assisted instructions (CAI) within the Cognitive Theory of Multimedia Learning: reducing extraneous processing (i.e., excluding irrelevant content), managing essential processing (i.e., focusing on the complex but essential learning materials), and fostering generative processing (i.e., maximizing learner motivation with multimedia features). No study so far has systematically evaluated each design feature, or their combinations for students with reading difficulties. The present study is the first meta-analysis to investigate the optimal design features of CAIs for students with reading difficulties on their decoding/word reading and reading comprehension performance. A total of 49 experimental studies were reviewed with Bayesian network meta-analysis. Results showed that CAI programs with features that reduced extraneous processing were the most effective (g = 0.65) followed by programs that combined reducing extraneous processing, managing essential processing, and fostering generative processing (g = 0.29) as well as programs that combined reducing extraneous processing and managing essential processing (g = 0.27). CAI programs yielded larger effects when they were designed for younger learners with reading difficulties compared to older learners. No significant moderation effects were observed for students' reading difficulty status, reading content, reading outcomes, instruction dosages, control group types, measures, and fidelity checks. These findings suggest that different combinations of design features of CAI programs may generate different effects. Lowering students' cognitive loads by excluding irrelevant content may be the foundation for designing effective computer instructions for students with reading difficulties.