Automatic variational inference (ADVI) can be implemented in Stan to automate variational inference for any probabilistic model specified in Stan. ADVI determines an appropriate variational family and optimizes the variational objective without any input from the user beyond providing the model and data. ADVI handles nonconjugate models by automatically deriving an inference algorithm. It scales to large datasets using subsampling and has been shown to outperform sampling methods while training models with hundreds of thousands of data points.