Gesture recognition using millimeter-wave(mmWave) radar offers advantages such as contact-free operation, high detection accuracy, privacy preservation, and robust environmental adaptability, making it promising for industrial human-machine interaction and smart-home applications. However, existing mmWave-based dynamic-gesture recognition approaches suffer from high model complexity, large computational cost, low accuracy, and slow inference speed. To address these challenges, a lightweight gesture-recognition method is proposed based on an improved MobileViT network that maintains high accuracy while significantly reducing computational complexity for deployment on embedded devices. Firstly, dynamic-gesture echoes are captured with an mmWave radar. After suppressing device noise and background clutter, the data are reorganized into a 3-D matrix(sample points×chirps×frames). Range-time and Doppler-time maps are then generated via Fourier transform and fed into the enhanced MobileViT model for feature extraction and fusion, yielding the final gesture classification. Experimental results show that the proposed MobileViT model has only 0.167 M of parameter space complexity and 0.253 GFLOPs of computational complexity. Evaluated on a 12-class dynamic-gesture dataset, the method achieves 99.31% of recognition accuracy, demonstrating its effectiveness.