Few-shot learning (FSL) has been extensively studied in the 2D domain, significantly propelling its development in the 3D field. Current 3D FSL methods predominantly rely on Prototypical Networks, which face challenges in acquiring high-quality representations, leading to considerable prototype bias. Although some approaches integrate 3D point cloud features with multi-view counterparts to achieve more discriminative representations, they often fail to address prototype bias issues such as missing points and occlusion, while also overlooking the semantic consistency between 3D point clouds and multi-view modalities. To tackle these challenges, we introduce a novel 3D FSL framework called Semantic-Guided Dual-branch Co-inference Network (SGDCN). SGDCN utilizes class-specific semantics to guide prototype reconstruction for both point cloud and multi-view modalities via the proposed PSA and MSA modules, effectively reducing prototype bias. Moreover, our dual-modality mutual distillation and ensemble strategies further enhance FSL performance. Extensive experiments on three well-known benchmarks demonstrate the efficacy of our proposed method. The source code is available at https://github.com/Hamlynn/SGDCN .