On-Device ML - What Does Low Latency Mean?

The model in the cloud, which would eventually cost money, depending on how much I use this thing, had more information. Therefore, it could get closer to the pictures that I was drawing with these doodles than I could on the device. The problem was if it only had 400 from a universe of infinity, then it wasn't going to be very effective. If it only had to detect, let's say, 26, but I could pre-train the model, then that probably would have been the best of both worlds. So for me, what low latency mean is actually new user experiences.