in what cases should I fine time and/or use the base model directly?

is it an alternative of embedding models like ada or sentence transformer embeddings?