Valproate is an old-generation antiepileptic drug often used in children. The pharmacokinetics of valproate are noteworthy for a large and difficult to predict interindividual variability in measured serum concentrations and for saturable protein binding. A model-based approach to personalize valproate treatment could be relevant in pediatric patients. The aims of this study were to review all published valproate population pharmacokinetic models in children and assess them by external validation to determine their predictive performance. Through simulations with the best model, we evaluated dosing regimen. A validation data set included valproate serum concentrations assayed during routine therapeutic drug monitoring of epileptic children. We applied to our population 11 published pediatric population pharmacokinetic models. For each model, predictive performance was assessed by external validation, using bias and precision calculations as well as goodness-of-fit plots. Dose simulations were conducted with the best predictive model to evaluate dosing regimen. The validation data set contained 178 valproate concentrations ranging from 13.4 to 128 mg/L from 114 patients. The best model exhibited a mean prediction error of 6.6 mg/L and a root mean squared error of 25.1 mg/L, with no model misspecification evidenced by visual predictive check. In our cohort, half the patients had a trough concentration <50 mg/L. Simulations suggested increasing doses, especially for children ࣘ40 kg. External evaluation of published valproate pharmacokinetic models enabled us to identify a suitable model for simulations and Bayesian forecasting. Dosing regimen should be adjusted to weight, with decreasing doses with increasing weight.