The paper introduces the Temporal Shift Module (TSM), a novel approach for efficient video understanding. TSM is designed to achieve both high performance and low computational cost, making it suitable for real-time applications on edge devices. The module shifts part of the channels along the temporal dimension, facilitating information exchange between neighboring frames without increasing computational complexity. TSM can be integrated into 2D CNNs to enhance temporal modeling while maintaining the same complexity as 2D CNNs. The paper also extends TSM to online settings, enabling real-time video recognition with low latency. Experimental results show that TSM outperforms existing methods on various datasets, achieving state-of-the-art performance on the Something-Something leaderboard and demonstrating low latency and high throughput on hardware tests. The code for TSM is available at <https://github.com/mit-han-lab/temporal-shift-module>.The paper introduces the Temporal Shift Module (TSM), a novel approach for efficient video understanding. TSM is designed to achieve both high performance and low computational cost, making it suitable for real-time applications on edge devices. The module shifts part of the channels along the temporal dimension, facilitating information exchange between neighboring frames without increasing computational complexity. TSM can be integrated into 2D CNNs to enhance temporal modeling while maintaining the same complexity as 2D CNNs. The paper also extends TSM to online settings, enabling real-time video recognition with low latency. Experimental results show that TSM outperforms existing methods on various datasets, achieving state-of-the-art performance on the Something-Something leaderboard and demonstrating low latency and high throughput on hardware tests. The code for TSM is available at <https://github.com/mit-han-lab/temporal-shift-module>.