Comment by tosh

It is both cost efficient in training (+ future fine-tuning) as well as inference compared to most other current models.

Can you elaborate?