Facial paralysis seriously affects patients’ quality of life, necessitating accurate severity assessment for effective treatment planning. Traditional House-Brackmann grading relies on subjective clinical evaluation, leading to low efficiency. Recent deep learning techniques offer the prospect of automatic facial paralysis assessment, but existing methods predominantly utilize 2D facial images, the accuracy is limited due to loss of spatial information. To address this, this paper presents the first study of automatic facial paralysis severity assessment using 3D point clouds. A novel facial paralysis dataset is collected, which comprises of 61 patients exhibiting eight clinically recommended facial expressions, resulting in 488 annotated samples across House-Brackmann grades II-V. We employ advanced 3D point clouds preprocessing techniques including landmark detection and spherical cropping. We evaluate SOTA 3D recognition models including PointNet++, PointMLP, and PointNN. Experimental results demonstrate the superiority of 3D-based assessment methods against 2D-based approaches. Specifically, PointNet++, PointMLP and PointNN respectively achieve assessment accuracies of 35.00%, 40.13% and 37.50%, significantly surpassing 32.00% accuracy obtained from 2D-based approach. Our findings exhibit robust baselines for 3D facial paralysis severity assessment and validate the clinical potential of point cloud-based 3D recognition models.