The article introduces i-PI 3.0, a flexible and efficient framework for advanced atomistic simulations. i-PI is designed to integrate machine-learning interatomic potentials (MLIPs) with advanced modeling techniques, leveraging a modular software architecture based on inter-process communication through sockets. The new release includes several optimizations and new features, such as efficient algorithms for modeling bosonic and fermionic exchange, uncertainty quantification, and integration with electronic-driven simulations. The authors benchmark and optimize i-PI for common simulation scenarios, making computational overhead negligible for systems with tens of thousands of atoms using popular MLIPs like Behler-Parinello, DeePMD, and MACE neural networks. The article also discusses the efficiency of large-scale simulations, demonstrating that the overhead introduced by i-PI is negligible or small up to high levels of parallelization of the MLIP code. Additionally, it highlights the implementation of advanced features like bosonic and fermionic path integral molecular dynamics, nuclear propagation by additional dynamical variables, and non-adiabatic tunnelling rates in metallic systems. The modular structure of i-PI allows for easy integration with various client codes, making it a versatile tool for a wide range of applications in atomic-scale modeling.The article introduces i-PI 3.0, a flexible and efficient framework for advanced atomistic simulations. i-PI is designed to integrate machine-learning interatomic potentials (MLIPs) with advanced modeling techniques, leveraging a modular software architecture based on inter-process communication through sockets. The new release includes several optimizations and new features, such as efficient algorithms for modeling bosonic and fermionic exchange, uncertainty quantification, and integration with electronic-driven simulations. The authors benchmark and optimize i-PI for common simulation scenarios, making computational overhead negligible for systems with tens of thousands of atoms using popular MLIPs like Behler-Parinello, DeePMD, and MACE neural networks. The article also discusses the efficiency of large-scale simulations, demonstrating that the overhead introduced by i-PI is negligible or small up to high levels of parallelization of the MLIP code. Additionally, it highlights the implementation of advanced features like bosonic and fermionic path integral molecular dynamics, nuclear propagation by additional dynamical variables, and non-adiabatic tunnelling rates in metallic systems. The modular structure of i-PI allows for easy integration with various client codes, making it a versatile tool for a wide range of applications in atomic-scale modeling.