RELION-3 is the third major release of the RELION software for automated high-resolution cryo-EM structure determination. It introduces several improvements, including CPU-based vector acceleration alongside GPU support, enhancing flexibility and reducing memory limitations. Reference-free autopicking using Laplacian-of-Gaussian filtering and Python job execution enable non-interactive processing during data acquisition, including 2D classification, de novo model generation, and 3D classification. Per-particle refinement of CTF parameters and beam tilt correction improve resolution for particles at different heights in the ice or with suboptimal coma-free alignment. Ewald sphere curvature correction enhances resolution for large particles. These developments are illustrated with publicly available datasets, leading to resolution improvements of 0.2–0.7 Å compared to previous versions. Key features include Laplacian-of-Gaussian auto-picking for particle selection, per-particle CTF refinement using a 3D reference structure, and beam-tilt estimation and correction. The Ewald sphere correction algorithm improves resolution by addressing curvature effects. CPU vector acceleration allows efficient processing on standard hardware, reducing memory requirements and enabling higher resolution reconstructions. Python scripts enable automated processing, with examples like bfactor_plot.py and relion_it.py for data quality feedback and standardised tasks. Additional developments include Bayesian particle polishing and multi-body refinement, enhancing resolution and providing insights into complex structures. RELION-3 also improves performance on x86 CPUs, achieving cost-efficiency comparable to professional GPUs. Per-particle defocus correction and beam tilt correction significantly improve resolution, as demonstrated by reprocessing datasets with tilted samples. These advancements make RELION-3 a powerful tool for high-resolution cryo-EM structure determination.RELION-3 is the third major release of the RELION software for automated high-resolution cryo-EM structure determination. It introduces several improvements, including CPU-based vector acceleration alongside GPU support, enhancing flexibility and reducing memory limitations. Reference-free autopicking using Laplacian-of-Gaussian filtering and Python job execution enable non-interactive processing during data acquisition, including 2D classification, de novo model generation, and 3D classification. Per-particle refinement of CTF parameters and beam tilt correction improve resolution for particles at different heights in the ice or with suboptimal coma-free alignment. Ewald sphere curvature correction enhances resolution for large particles. These developments are illustrated with publicly available datasets, leading to resolution improvements of 0.2–0.7 Å compared to previous versions. Key features include Laplacian-of-Gaussian auto-picking for particle selection, per-particle CTF refinement using a 3D reference structure, and beam-tilt estimation and correction. The Ewald sphere correction algorithm improves resolution by addressing curvature effects. CPU vector acceleration allows efficient processing on standard hardware, reducing memory requirements and enabling higher resolution reconstructions. Python scripts enable automated processing, with examples like bfactor_plot.py and relion_it.py for data quality feedback and standardised tasks. Additional developments include Bayesian particle polishing and multi-body refinement, enhancing resolution and providing insights into complex structures. RELION-3 also improves performance on x86 CPUs, achieving cost-efficiency comparable to professional GPUs. Per-particle defocus correction and beam tilt correction significantly improve resolution, as demonstrated by reprocessing datasets with tilted samples. These advancements make RELION-3 a powerful tool for high-resolution cryo-EM structure determination.