Deploying efficient deep learning models in resource-constrained environments is challenging due to their growing scale and complexity. Knowledge Distillation (KD) offers a practical solution by transferring knowledge from a large teacher model to a compact student model. However, traditional KD methods often fall short when significant capacity gaps exist. To address this issue, we introduce Gap-KD, which utilizes dynamic temperature scaling and a double decoupling technique to bridge these gaps. The distillation temperature is adjusted dynamically, progressively increasing the student model’s learning difficulty. A Teacher Assistant (TA) model is introduced as an intermediary layer, initially reducing the gap. Building on this, the outputs of the teacher, TA, and student models are doubly decoupled, further reducing information loss and error accumulation. Extensive experiments on CIFAR-10 and CIFAR-100 datasets with ResNet and CNN architectures demonstrate that Gap-KD achieves state-of-the-art performance specifically in scenarios with significant capacity gaps, highlighting its effectiveness for these challenging conditions. The code is available at https://github.com/kennysH/Gap-KD .