[abdinoor]

(Full disclosure: I work at Fritz, and we publish Heartbeat).

There are a couple of legitimate reasons that the on-device model is lower accuracy: compute and size. Models are often reduced in complexity when running on edge devices, but cloud-based GPU (TPU in Google's case) models can require far more calculations in order to squeeze that bit of accuracy. As for model size, a 100MB model in the cloud is not a big deal, but having to download a 100MB model on every edge device is expensive in both time and bandwidth.

Outside of these reasons, Google/Firebase may want on-device capabilities limited in order push adoption of their cloud ML services.

[Razengan]

> compute and size ... having to download a 100MB model on every edge device is expensive in both time and bandwidth.

The WWDC 2018 ML sessions showed some solutions to address these issues. (I don't exactly recall which as I'm not using ML yet, so I skimmed through them.)