Millimeter-wave (mmWave) radar provides robust and privacy preserving sensing for indoor object detection on edge devices, which is critical for smart environments and healthcare applications. However, acquiring numerous labeled data for training neural networks is often prohibitively expensive in practice. Furthermore, conventional self-supervised and semi-supervised methods offer limited effectiveness in leveraging unlabeled data when applied to lightweight models. To address these challenges, we propose an Active Semi-Supervised Learning (ASSL) framework for mmWave radar target classification in label-scarce scenarios. ASSL actively selects a small subset of informative samples for annotation to further guide model effectively learning from unlabeled data, tailored for lightweight networks. Central to this framework is a novel Hybrid Uncertainty-Diversity Query (HUDQ) strategy, which combines model uncertainty (identifying ambiguous samples) with data diversity (ensuring broad coverage) to enhance model generalization. By balancing these complementary criteria, ASSL effectively identifies representative samples for labeling, following with a robust semi-supervised learning scheme to train a lightweight CNN classifier. Experimental results on a real-world mmWave radar dataset demonstrate that ASSL, using only 20% labeled data, surpasses the performance of fully supervised models on a lightweight architecture. This highlights its practical value for AIoT applications on edge devices with lightweight models and limited annotated data.