The problem of cognitive overload is particularly pertinent in multimedia L2 classroom corrective feedback (CF), which involves rich communicative tools to help the class to notice the mismatch between the target input and learners' pronunciation. Based on multimedia design principles, this study developed a new multimodal CF model through augmented reality (AR)-enabled videoconferencing to eliminate extraneous cognitive load and guide learners' attention to the essential material. Using a quasi-experimental design, this study aims to examine the effectiveness of this new CF model in improving Chinese L2 students' segmental production and identification of the targeted English consonants (dark /[voiced velar lateral approximant]/, /[voiced dental fricative]/and /[voiceless dental fricative]/), as well as their attitudes towards this application. Results indicated that the online multimodal CF environment equipped with AR annotation and filters played a significant role in improving the participants' production of the target segments. However, this advantage was not found in the auditory identification tests compared to the offline CF multimedia class. In addition, the learners reported that the new CF model helped to direct their attention to the articulatory gestures of the student being corrected, and enhance the class efficiency. Implications for computer-assisted pronunciation training and the construction of online/offline multimedia learning environments are also discussed.