<p>Accurate prediction of biomolecular complex structures is fundamental for understanding biological processes and rational therapeutic design. Recent advances in deep learning methods, particularly all-atom structure prediction models, have significantly expanded their capabilities to include diverse biomolecular entities, such as proteins, nucleic acids, ligands, and ions. However, comprehensive benchmarks covering multiple interaction types and molecular diversity remain scarce, limiting fair and rigorous assessment of model performance and generalizability. To address this gap, we introduce FoldBench, an extensive benchmark dataset consisting of 1522 biological assemblies categorized into nine distinct prediction tasks. Our evaluations reveal critical performance dependencies, showing that ligand docking accuracy notably diminishes as ligand similarity to the training set decreases, a pattern similarly observed in protein-protein interaction modeling. Furthermore, antibody-antigen predictions remain particularly challenging, with current methods exhibiting failure rates exceeding 50%. Among evaluated models, AlphaFold 3 consistently demonstrates superior accuracy across the majority of tasks. In summary, our results highlight significant advancements yet reveal persistent limitations within the field, providing crucial insights and benchmarks to inform future model development and refinement.</p>