<p>The Versatile Video Coding (VVC) standard achieves approximately 50% bitrate savings over its predecessor, High Efficiency Video Coding (HEVC), through a series of advanced coding tools. However, its increased computational complexity, particularly the exhaustive decision-making process associated with the quadtree with nested multi-type tree (QTMT) partitioning, severely hinders real-time applications. To address this challenge, this paper proposes an efficient QTMT partitioning approach for VVC, based on a lightweight parameter-conditioned dynamic routing network (PDR-Net). Specifically, PDR-Net comprises three key components: the parameter fusion unit, the gated multiscale (GMS) unit, and the dynamic routing unit. The parameter fusion unit integrates CU’s quantization parameters and partitioning constraints to enable accurate partitioning across varying distortion levels, while ensuring that only valid partition modes are generated. The GMS unit is specifically designed for effective feature extraction and enhancement on small-size CUs through parallel attention branches and a gated residual network. The dynamic routing unit further improves the network’s adaptability to diverse content characteristics through learnable routing weights. To train PDR-Net, we propose a Monte Carlo-based probability labeling method, which generates a probability distribution over partitioning modes for each CU to supervise the training process. Experimental results on the VVC reference software (VTM-15.0) demonstrate that our method achieves 40.71% to 62.41% time reduction with only 0.34% to 1.30% Bjøntegaard Delta Bit-Rate (BD-BR) increase, significantly outperforming existing approaches. Our model size is only 706 KB, which is attractive to practical for deployment on resource-constrained multimedia devices.</p>