Objective tools for the assessment of animal welfare are needed. The present study analyzed the interobserver reliability of the Animal Welfare Indicators (AWIN) welfare assessment protocol for horses to further enhance knowledge concerning reliability. Therefore, two trained observers conducted 18 assessments on farm at the same time and on the same animals. The results were compared at individual level by calculation of Cohen's kappa (κ), weighted kappa (κw), and prevalence-adjusted, bias-adjusted kappa (PABAK). Spearman rank correlation coefficient (RS), intraclass correlation coefficient (ICC), limits of agreement (LoA), and smallest detectable change (SDC) were used at farm level. The Qualitative Behaviour Assessment was further analyzed by means of principal component analysis. At the individual level, most of the indicators demonstrated acceptable (κ, κw, PABAK ≥ 0.4) to good (κ, κw, PABAK ≥ 0.6) interobserver reliability. Also, at farm level, most of the indicators demonstrated acceptable (RS ≥ 0.4; ICC ≥ 0.4; SDC: ≤ 0.1; LoA ε [0.1; 0.1]) to good (RS: ≥ 0.7; ICC: ≥ 0.7; SDC: ≤ 0.05; LoA: ε [0.05; 0.05]) interobserver reliability. Exceptions were the indicators moderate presence of tension above eye area (score 1) and orbital tightening (score 1) on the Horse Grimace Scale, as well as the presence of swollen joints. Furthermore, the present results indicate that the details for the different scores should be improved for some indicators such as the Body Condition Score. In general, this study points out a good interobserver reliability of the AWIN welfare assessment protocol for horses.