They're probably referring to fine-tuning on private/proprietary data that is specific to a use case. Say a history of conversation transcripts in a call center.

Larger models, like OpenAI's GPT, don't have access to this by default.