This paper systematically evaluates four different universal machine-learning interatomic potentials (uMLIPs) based on graph neural network (GNN) architectures to assess their performance in materials modeling. The evaluation uses datasets from density-functional-theory (DFT) implementations and the Materials Project (MP). The four uMLIPs evaluated are M3GNet-DIRECT, CHGNet, MACE-MP-0, and ALIGNN-FF. The study tests the models' capabilities in predicting the equation of state, formation energies, structural optimizations, and vibrational properties. Key findings include: 1. **Equation of State**: MACE-MP-0 and CHGNet show superior performance in predicting the equation of state compared to other uMLIPs. 2. **Formation Energies**: MACE-MP-0 and CHGNet perform best in predicting formation energies, especially for transition metals, chalcogens, and halogens. 3. **Structural Optimization**: CHGNet and M3GNet are more robust in performing structural optimizations, with lower fractions of unconverged results. 4. **Vibrational Properties**: MACE-MP-0 and CHGNet show the best performance in predicting vibrational properties, with MACE-MP-0 having the smallest mean absolute error (MAE) in phonon band structures. The study concludes that while uMLIPs have shown promise, further optimization and training are needed to enhance their accuracy and applicability across a broader range of materials and properties. The choice of uMLIP should consider a balance between accuracy and computational efficiency, and future work should focus on improving the prediction of forces, stresses, and the ability to learn Born effective charges and electronic dielectric tensors for polar materials.This paper systematically evaluates four different universal machine-learning interatomic potentials (uMLIPs) based on graph neural network (GNN) architectures to assess their performance in materials modeling. The evaluation uses datasets from density-functional-theory (DFT) implementations and the Materials Project (MP). The four uMLIPs evaluated are M3GNet-DIRECT, CHGNet, MACE-MP-0, and ALIGNN-FF. The study tests the models' capabilities in predicting the equation of state, formation energies, structural optimizations, and vibrational properties. Key findings include: 1. **Equation of State**: MACE-MP-0 and CHGNet show superior performance in predicting the equation of state compared to other uMLIPs. 2. **Formation Energies**: MACE-MP-0 and CHGNet perform best in predicting formation energies, especially for transition metals, chalcogens, and halogens. 3. **Structural Optimization**: CHGNet and M3GNet are more robust in performing structural optimizations, with lower fractions of unconverged results. 4. **Vibrational Properties**: MACE-MP-0 and CHGNet show the best performance in predicting vibrational properties, with MACE-MP-0 having the smallest mean absolute error (MAE) in phonon band structures. The study concludes that while uMLIPs have shown promise, further optimization and training are needed to enhance their accuracy and applicability across a broader range of materials and properties. The choice of uMLIP should consider a balance between accuracy and computational efficiency, and future work should focus on improving the prediction of forces, stresses, and the ability to learn Born effective charges and electronic dielectric tensors for polar materials.