Multimodal Emotion Recognition (MER) models have made significant progress with the development of deep learning, and most of them are deployed on remote servers. With the gradual popularization and widespread use of mobile terminal devices, local deployment has become more valuable in practical applications. However, current MER models often contain complex neural network structures, resulting in high computational costs and challenges in local deployment. Moreover, the current lightweight MER models post limited accuracy. To address these issues, we proposed a curriculum multimodal knowledge distillation framework (MKDF), which includes feature-based distillation and curriculum temperature optimisation, with frame-by-frame video extraction, aiming to improve performance while keeping lightweight model parameters. For model training and validation, we use the IEMOCAP and CMU-MOSEI multimodal emotion datasets and achieve an accuracy of 67.26% and 49.26%, which is close to state-of-the-art methods while keeping 1.69M parameters only, which is much smaller than other approaches. The source code are available at https://github.com/ItsDia/MKDF