Google kept promoting the "speed" of the model, so I guess it will be useful for some close to real-time use-cases, maybe live chat/support (?)