Deep neural networks have demonstrated exceptional efficacy in image classification domains. However, their performance is significantly degraded when trained on long-tailed distributed data. To tackle this challenge, numerous researchers have explored MoE-based (Mixture of Experts) approaches, which specialize in different segments of the long-tailed distribution. Despite their potential, existing methods typically employ experts with uniform depth, limiting knowledge transfer to logit-level information in the output layer while overlooking the valuable “latent dark knowledge” embedded in intermediate layers. This work introduces a new method called Attention-guided Feature Distillation for Long-Tailed Visual Recognition (AFDLT), implemented within the framework of MoE. First, we introduce depth-wise knowledge fusion, which implicitly assigns different preferences from the long-tailed distribution to experts, enhancing the diversity of expert features. Then, we incorporate an attention mechanism to highlight key features beneficial for distinguishing images. Using attention feature maps as the distillation target, we enhance the recognition ability of tail classes through feature knowledge transfer between experts. Comprehensive evaluations across standard long-tailed benchmarks (CIFAR10/100-LT, ImageNet-LT) and real-world scenarios (iNaturalist2018) demonstrate the framework’s efficacy.