Twenty-two loci for soybean SW and candidate genes conditioning seed development were identified; and prediction accuracies of GS and MAS were estimated through cross-validation and validation with unrelated populations. Soybean (Glycine max) is a major crop for plant protein and oil production, and seed weight (SW) is important for yield and quality in food/vegetable uses of soybean. However, our knowledge of genes controlling SW remains limited. To better understand the molecular mechanism underlying the trait and explore marker-based breeding approaches, we conducted a genome-wide association study in a population of 309 soybean germplasm accessions using 31,045 single nucleotide polymorphisms (SNPs), and estimated the prediction accuracy of genomic selection (GS) and marker-assisted selection (MAS) for SW. Twenty-two loci of minor effect associated with SW were identified, including hotspots on Gm04 and Gm19. The mixed model containing these loci explained 83.4% of phenotypic variation. Candidate genes with Arabidopsis orthologs conditioning SW were also proposed. The prediction accuracies of GS and MAS by cross-validation were 0.75-0.87 and 0.62-0.75, respectively, depending on the number of SNPs used and the size of training population. GS also outperformed MAS when the validation was performed using unrelated panels across a wide range of maturities, with an average prediction accuracy of 0.74 versus 0.53. This study convincingly demonstrated that soybean SW is controlled by numerous minor-effect loci. It greatly enhances our understanding of the genetic basis of SW in soybean and facilitates the identification of genes controlling the trait. It also suggests that GS holds promise for accelerating soybean breeding progress. The results are helpful for genetic improvement and genomic prediction of yield in soybean.