https://skymel.com/ . We are in closed beta and provide javascript library ( along with api endpoint) that decides in real time whether to run inference locally on the device or call api . You only pay for the api calls or gpu minutes used.