3 ms·

First reason that comes to mind is you can make much smaller models, which helps with latency, cost and may enable you to run the model locally.