<p>Infrared small-target detection (ISTD) aims to accurately segment dim and compact targets from infrared images under complex background interference. However, deploying ISTD models in resource-limited edge environments imposes strict constraints on computational complexity, model size, and inference latency. Although Mamba-based models provide an efficient solution for modeling long-range contextual dependencies, converting two-dimensional feature maps into one-dimensional sequences may weaken local neighborhood information that is essential for small-target localization. To address this limitation, we propose the Physics-Guided Frequency-Spatial Mamba Network (PFSMNet), a lightweight model for real-time ISTD. PFSMNet first uses the Multi-Scale Top-Hat Attention Module to enhance compact bright structures and the Wavelet Kolmogorov–Arnold Network to learn adaptive wavelet priors. These priors are then introduced into Physics-Guided Mamba before selective scanning, enabling frequency information to guide Mamba modeling. In addition, the Hybrid Frequency-Spatial Bridge refines multi-scale skip features, and the Gradient-Aware Decoding Mechanism improves boundary reconstruction. Extensive experiments on three public benchmarks demonstrate that PFSMNet achieves an effective balance between detection accuracy and computational efficiency. The model contains only 0.0529 million parameters and requires 0.3140 G floating-point operations. Deployed on NVIDIA Jetson Orin NX, PFSMNet achieves 65.36 frames per second, further confirming its potential for real-time edge inference.</p>