A smaller model can respond faster, run closer to the user, and process high-volume tasks at a lower cost. For narrow workflows, careful examples, retrieval, and evaluation may matter more than general model scale.
What matters in practice
Local or private deployment can simplify data-governance requirements. It also creates operational work: model distribution, hardware capacity, monitoring, and a plan for upgrades all become part of the product.
A useful rule
Measure on representative tasks. Track correctness, latency, cost, refusal behavior, and failure patterns. A model is useful when the complete system performs reliably for the people using it.