This paper presents the technical development and accuracy assessment of the SoilGrids system at 250m resolution, an improved version of the original SoilGrids system released in 2014. The updated SoilGrids provides global predictions for standard numeric soil properties (organic carbon, bulk density, Cation Exchange Capacity (CEC), pH, soil texture fractions, and coarse fragments) at seven standard depths (0, 5, 15, 30, 60, 100, and 200 cm), as well as depth to bedrock and soil class distribution based on the World Reference Base (WRB) and USDA classification systems. The predictions are based on approximately 150,000 soil profiles and a stack of 158 remote sensing-based soil covariates, including MODIS land products, SRTM DEM derivatives, climatic images, and global landform and lithology maps. Machine learning methods, such as random forest, gradient boosting, and multinomial logistic regression, were used to fit the models, achieving an overall average of 61% of variation explained. The improvements over the previous 1km resolution version include the use of machine learning, finer-resolution covariate layers, and additional soil profiles. The paper also discusses future improvements, such as incorporating input uncertainties and generating posterior probability distributions, and suggests ways to enhance the system, including increasing the quality and quantity of training data and combining global predictions with local prediction models.This paper presents the technical development and accuracy assessment of the SoilGrids system at 250m resolution, an improved version of the original SoilGrids system released in 2014. The updated SoilGrids provides global predictions for standard numeric soil properties (organic carbon, bulk density, Cation Exchange Capacity (CEC), pH, soil texture fractions, and coarse fragments) at seven standard depths (0, 5, 15, 30, 60, 100, and 200 cm), as well as depth to bedrock and soil class distribution based on the World Reference Base (WRB) and USDA classification systems. The predictions are based on approximately 150,000 soil profiles and a stack of 158 remote sensing-based soil covariates, including MODIS land products, SRTM DEM derivatives, climatic images, and global landform and lithology maps. Machine learning methods, such as random forest, gradient boosting, and multinomial logistic regression, were used to fit the models, achieving an overall average of 61% of variation explained. The improvements over the previous 1km resolution version include the use of machine learning, finer-resolution covariate layers, and additional soil profiles. The paper also discusses future improvements, such as incorporating input uncertainties and generating posterior probability distributions, and suggests ways to enhance the system, including increasing the quality and quantity of training data and combining global predictions with local prediction models.