Simulink C code generation vs MEX files

When generating C code from a Simulink model containing normal-mode references to other models, I can see that MEX files are created in the process for those other models.

As far as I understand, those MEX files are usually created when compiling a model which references other models in accelerator mode. Therefore I don't understand the need to create them when generating C code for a parent model having no accelerator-mode references.

Can anyone explain why this is necessary? If not, how can I bypass the creation of those MEX files when generating code?