The action recognition in the video expands broadly as a technically interesting and important task of computer vision across a wide variety of areas such as security, medical, and sport monitoring. The modeling result of the task in hand is effective when both space and time modeling are done effectively. In an effort to resolve these challenges, we put forth a new multi-model ensemble framework which uses four different types of deep learning networks: MovNet-LSTM, Long-Term Recurrent Convolutional Network (LRCN), 3D Convolutional Neural Network (CNN) and EfficientNet. MovNet-LSTM is capable of capturing space information and constructing a temporal model. On the other hand, LRCN uses convolutional and pooling layers to construct adequate spatial information first and only then learns a temporal model. 3D CNN uses depth filters to capture the spatial and motion information of the data set, and the model based on EfficientNet uses transfer learning to improve the model. By taking the outputs of the models and combining them through averaging, a set model was able to reach a high accuracy of 94.35% on the very challenging UCF-101 dataset while also outperforming individual models and demonstrating strength in handling complex video conditions such as fast movements, overlapping objects, and dynamic backgrounds. The proposed framework is computationally efficient and using only 18.1 million parameters, making it suitable for real-world applications. Our results highlight the effectiveness of ensemble methods in video action recognition and their potential to address the limitations of single-model approaches.