Abstract: Current detection methods are less effective for hateful meme with implicit semantics and culture-specific content. This study developed HMUM, a hateful meme understanding model based on Qwen2.5-VL-72B-Instruct. The model employed LoRA fine-tuning and a multi-stage prompt learning framework. It enhanced detection performance through progressive analysis of text recognition, emotion modeling, and hate reasoning. Evaluation on the ToxiCN MM dataset showed that HMUM achieved an 81.7% F₁-score for overall hate meme detection, i.e. an 8.4% improvement over the MKE baseline. For the implicit hate meme subset, the model reached 70.6% recall and 82.8% F₁-score.