The goal of few-shot object detection (FSOD) is to detect novel categories with limited labelled instances available for training. The recently emerging vision-language models (VLMs) bring significant improvements to few-shot learning tasks because they are capable of aligning visual embeddings with language embeddings. However, existing FSOD models only focus on VLMs’ text knowledge, while their visual knowledge can be further explored to help learn novel semantic concepts. Inspired by recent advancements in knowledge distillation methods using VLMs, we propose a feW-shOt Object Detector WIth kNowledge Distillation, termed WOODWIND. It utilizes a proposed Decoupling Knowledge Distillation Module to inject visual concepts from VLMs to few-shot object detection models. We also propose a simple yet effective online proposal sampling strategy to generate useful region proposals for the knowledge distillation head. Besides, a fine-tuning mechanism with knowledge distillation is devised to alleviate the over-fitting problems in the fine-tuning stage and help FSOD detector generalize to novel classes. Our proposed WOODWIND obtained state-of-the-art performance on both MS COCO and PASCAL VOC datasets on most of the few-shot settings.