MatterSim is a deep learning model that enables efficient atomistic simulations at the first-principles level and accurate prediction of a wide range of material properties across the periodic table, spanning temperatures from 0 to 5000 K and pressures up to 1000 GPa. The model is trained using large-scale first-principles computations and serves as a machine learning force field, capable of predicting ground-state material structures and energetics, as well as simulating their behavior under realistic temperatures and pressures. It achieves a ten-fold increase in accuracy compared to previous methods, enabling the computation of lattice dynamics, mechanical and thermodynamic properties with accuracy comparable to first-principles methods. MatterSim also serves as a platform for continuous learning and customization by integrating domain-specific data, allowing it to be fine-tuned for atomistic simulations at a desired level of theory or for direct structure-to-property predictions, achieving high data efficiency with a reduction in data requirements by up to 97%. MatterSim employs an active learning approach to explore the extensive materials space, integrating a deep graph neural network, a materials explorer, a first-principles supervisor, and an ensemble uncertainty monitor. The model is trained on a large dataset of materials structures, covering a wide range of temperatures and pressures, and is capable of learning a wide range of materials space with minimal data redundancy. The model's extensive coverage of the compositional and configurational space of materials enables it to effectively describe material features in the latent space and to serve as a pre-trained model for continuous learning and further customization, with high data efficiency. MatterSim is capable of predicting temperature- and pressure-dependent free energies of wide ranges of solid materials comparable with first-principles methods and experimental measurements, thereby opening an opportunity for fast and accurate prediction of phase diagrams of materials. The model can be extended to carry out atomistic simulations of highly complex systems beyond its current data coverage and theory level. For example, to simulate liquid water, only 3% of the data is needed to customize MatterSim to obtain the results of a specialized model trained from scratch, and to reproduce the experimental structural and transport properties of water. MatterSim's high expressive features enable direct structure-to-property prediction of materials, which is also known as end-to-end prediction. After being fine-tuned with a limited amount of data, MatterSim outperforms specialized models trained exclusively with domain-specific data on the tasks related to lattice dynamics, electronic and mechanical properties in Matbench. The model's performance is benchmarked on multiple open datasets as well as three newly created ones with better representation of the model's capability under finite temperatures and pressures. The results show that MatterSim achieves a sub-10 meV/atom error for temperatures up to 1000 K when compared with QHA computations at PBE level of theory, signifying a near-first-principles predictive power. MatterSim is also capable of performingMatterSim is a deep learning model that enables efficient atomistic simulations at the first-principles level and accurate prediction of a wide range of material properties across the periodic table, spanning temperatures from 0 to 5000 K and pressures up to 1000 GPa. The model is trained using large-scale first-principles computations and serves as a machine learning force field, capable of predicting ground-state material structures and energetics, as well as simulating their behavior under realistic temperatures and pressures. It achieves a ten-fold increase in accuracy compared to previous methods, enabling the computation of lattice dynamics, mechanical and thermodynamic properties with accuracy comparable to first-principles methods. MatterSim also serves as a platform for continuous learning and customization by integrating domain-specific data, allowing it to be fine-tuned for atomistic simulations at a desired level of theory or for direct structure-to-property predictions, achieving high data efficiency with a reduction in data requirements by up to 97%. MatterSim employs an active learning approach to explore the extensive materials space, integrating a deep graph neural network, a materials explorer, a first-principles supervisor, and an ensemble uncertainty monitor. The model is trained on a large dataset of materials structures, covering a wide range of temperatures and pressures, and is capable of learning a wide range of materials space with minimal data redundancy. The model's extensive coverage of the compositional and configurational space of materials enables it to effectively describe material features in the latent space and to serve as a pre-trained model for continuous learning and further customization, with high data efficiency. MatterSim is capable of predicting temperature- and pressure-dependent free energies of wide ranges of solid materials comparable with first-principles methods and experimental measurements, thereby opening an opportunity for fast and accurate prediction of phase diagrams of materials. The model can be extended to carry out atomistic simulations of highly complex systems beyond its current data coverage and theory level. For example, to simulate liquid water, only 3% of the data is needed to customize MatterSim to obtain the results of a specialized model trained from scratch, and to reproduce the experimental structural and transport properties of water. MatterSim's high expressive features enable direct structure-to-property prediction of materials, which is also known as end-to-end prediction. After being fine-tuned with a limited amount of data, MatterSim outperforms specialized models trained exclusively with domain-specific data on the tasks related to lattice dynamics, electronic and mechanical properties in Matbench. The model's performance is benchmarked on multiple open datasets as well as three newly created ones with better representation of the model's capability under finite temperatures and pressures. The results show that MatterSim achieves a sub-10 meV/atom error for temperatures up to 1000 K when compared with QHA computations at PBE level of theory, signifying a near-first-principles predictive power. MatterSim is also capable of performing