Ransomware has become one of the most damaging cyber threats in recent years. Unlike traditional detection methods that rely on manually engineered features, we propose a novel architecture, GLOW, which integrates semantic understanding into ransomware detection. GLOW is equipped with sliding window attention and a hierarchical and lightweight design to analyze the attacker’s intent from a semantic perspective. It achieves three major breakthroughs: capturing long-range behavioral dependencies, enhancing inference efficiency through hierarchical sparsity, and analyzing semantic relationships rather than relying on handcrafted feature engineering. Experiments demonstrate that GLOW achieves a detection precision of 98.2% and a false positive rate of only 2.5%, offering a promising direction for accurate and resilient ransomware detection.