問題文
Twelve business units each need their own adapted behavior, and the platform must serve all twelve from limited hardware with one deployment. Which design meets that?
選択肢
- Serve one merged model and switch the instruction text per unit, so that the twelve behaviors come from wording rather than from separate weights.
- Deploy twelve separate services, one per unit, each of them holding its own full copy of the weights on the same shared hardware as the rest.
- Keep one copy of the base weights resident and load the small per-unit differences alongside it, selecting the difference per request.
- Merge each unit's difference into its own copy of the base weights and load the twelve merged models; a merged model needs no extra work at inference time, and the twelve copies give the lowest possible latency for every unit.