The growing demand for drone-based fire detection drives the need for more efficient, lightweight algorithms. However, challenges remain in detecting small flame targets and enabling deployment on edge devices. This study proposes EAL-YOLO (Enhanced Attention Lightweight YOLO), a lightweight fire detection model based on YOLOv11n. Our approach integrates the Multi-Scale Feature Enhancement Attention Module (MSDA) with the C2PSA module to improve multi-scale feature extraction with minimal computational overhead. We replace the YOLOv11n backbone with the EfficientNet-v2s network, reducing complexity without sacrificing accuracy. Finally, we enhance feature fusion by introducing a Multi-Level Feature Fusion Network (MCFN) into the neck network’s C3k2 module. Experiments on a custom fire dataset show EAL-YOLO achieves a 77.9% mAP, representing a 3.1% improvement over baseline YOLOv11n, while significantly reducing parameters and computational load. Compared to existing models, EAL-YOLO improves accuracy while maintaining a lightweight design, making it highly suitable for edge platforms like drones.