Global kilometer-scale (km-scale) weather and climate models offer new opportunities to unify numerical weather prediction (NWP) and climate modeling by explicitly simulating convection and mesoscale circulations globally within a single modeling framework. We present results from the first multi-year (April 2020–March 2024) global atmosphere-land simulation using the GPU-refactored ICON model at a 2.5 km horizontal grid spacing and 120 vertical levels. The simulation uses NWP physics and observed sea-surface temperatures. We assess its performance against satellite, reanalysis, and in-situ observations using standard statistics and the MOAAP feature-tracking framework to evaluate a wide spectrum of atmospheric phenomena. ICON reproduces global temperature and precipitation patterns, including a realistic single Intertropical Convergence Zone and physically consistent diurnal precipitation cycles. However, ICON exhibits continental warm and dry biases during the warm season, linked to an overestimation of incoming solar radiation and excessive surface sensible heat fluxes. The model realistically captures the intensity and frequency of hourly precipitation and near-surface winds, as well as the structure and occurrence of tropical cyclones. Mesoscale convective systems (MCSs) exhibit realistic spatial initiation patterns, but their frequency is underestimated over oceans and overestimated over tropical land. Long-lived MCSs are too infrequent and small, while excess rainfall from shallow and mid-level clouds suggests overactive warm-cloud microphysics. These biases likely stem in part from an underrepresentation of convectively coupled equatorial waves. Our results demonstrate the feasibility and scientific value of multi-year global convection-permitting simulations for exploring the weather–climate system and local-scale extreme events, while identifying key directions for future model development.