Model performance when using a binary file

Hello, I’m new to that field and I was wondering if the compiled version of a model specified in Stan runs faster than when running the stan code itself. The time series package Prophet runs really fast and seems to translate the Stan code to C++ and compile it, is that the main reason for model training performance?

There’s actually no way to run Stan code without compiling it (translating to C++ and then compiling the C++), so this is what every interface to Stan does. Sometimes the code is compiled in advance, other times it’s compiled right before running it, but in all cases when running Stan you’re running compiled C++ code.