Interest in local models rarely comes from technical curiosity. It usually stems from a data-protection or contractual constraint.
The first question is always data classification: what genuinely cannot leave the organisation. Often that is a fraction of the content, and a hybrid setup is enough.
The second is hardware. A mid-sized model runs on a well-specified workstation or server today, but concurrent users push requirements up quickly.
The third is maintenance: model updates, evaluation and debugging all become internal responsibilities. Plan for that up front.