Running a model directly on a user's device (phone, laptop, etc.) rather than sending data to a remote server, which improves privacy and reduces latency.