Gesture recognition, an important direction of human-computer interaction, has drawn considerable attention for its natural and intuitive nature. However, existing mmWave radar-based methods suffer from complex preprocessing, heavy models, and limited robustness in multi-gesture scenarios. This paper proposes a lightweight neural network based on mmWave radar point clouds. Using 6,000 samples collected from five volunteers with a TI IWR1443 radar, a multi-feature fusion Transformer-TCN (TT-Net) is designed for spatiotemporal learning, achieving 97.57% accuracy. A progressive compression framework combining sparse pruning, knowledge distillation, and dynamic quantization further reduces model size and parameters by 30.52% and 40.80%, respectively, with only 0.14% accuracy loss. The resulting model supports efficient embedded deployment for practical gesture recognition.