Understanding and predicting the structure–property relationships in magnetic clusters remain challenging due to the complex interplay between geometry, electronic structure and spin configurations: density functional theory suffers from prohibitive computational cost for large-scale exploration, while existing graph neural networks struggle to balance long-range magnetic correlation capture, topological fidelity and scalability. We present MagCluster-GMAF, a physics-inspired hybrid graph architecture with magnetic cluster-aware serialization to preserve topological and electronic similarity during graph-to-sequence transformation, paired with a dual-branch design combining a Mamba-2 backbone for efficient long-range spin correlation modeling and a lightweight attention module as topological compensation. On a 2D doped magnetic material dataset, it achieved a total magnetic moment prediction MAE of 0.2683 μB, reducing the prediction error by 9.5% relative to the GPS + Transformer baseline and outperforming Graph-Mamba; at the 1400-node scale, it cut the GPU memory consumption by over 70% versus standard graph transformers while exhibiting near-linear empirical scaling at the studied graph scales. SHAP-based feature attribution analysis further identifies the key physicochemical descriptors governing the predictions, providing an effective approach for nanoscale magnetic material screening and the high-throughput screening of spintronic materials.