GADGET-2 is a new massively parallel TreeSPH code for cosmological simulations, capable of simulating collisionless dark matter with the N-body method and ideal gas with smoothed particle hydrodynamics (SPH). It conserves energy and entropy in regions without dissipation and allows adaptive smoothing lengths. Gravitational forces are computed using a hierarchical multipole expansion, optionally combined with a TreePM algorithm for long-range forces. Time integration uses a quasi-symplectic scheme with different timesteps for long- and short-range forces. The code is efficient in memory and communication bandwidth, and has been used for simulations with over 10^10 dark matter particles and over 250 million particles for SPH simulations with radiative cooling and star formation. It has been tested on various problems and is publicly available. The code uses a space-filling curve for domain decomposition, resulting in high flexibility and tree force errors independent of domain cutting. It is efficient and accurate, with a hierarchical tree algorithm for gravitational interactions and SPH for gas dynamics. The code handles periodic boundaries and uses a TreePM method for hybrid gravitational force computation. It has been used to study structure formation, dark matter, and other cosmological phenomena. The code is flexible, efficient, and accurate, with a focus on high dynamic range and high-resolution simulations.GADGET-2 is a new massively parallel TreeSPH code for cosmological simulations, capable of simulating collisionless dark matter with the N-body method and ideal gas with smoothed particle hydrodynamics (SPH). It conserves energy and entropy in regions without dissipation and allows adaptive smoothing lengths. Gravitational forces are computed using a hierarchical multipole expansion, optionally combined with a TreePM algorithm for long-range forces. Time integration uses a quasi-symplectic scheme with different timesteps for long- and short-range forces. The code is efficient in memory and communication bandwidth, and has been used for simulations with over 10^10 dark matter particles and over 250 million particles for SPH simulations with radiative cooling and star formation. It has been tested on various problems and is publicly available. The code uses a space-filling curve for domain decomposition, resulting in high flexibility and tree force errors independent of domain cutting. It is efficient and accurate, with a hierarchical tree algorithm for gravitational interactions and SPH for gas dynamics. The code handles periodic boundaries and uses a TreePM method for hybrid gravitational force computation. It has been used to study structure formation, dark matter, and other cosmological phenomena. The code is flexible, efficient, and accurate, with a focus on high dynamic range and high-resolution simulations.