Offline dictation is a property of the engine, not of the hotkey. A local model files itself into application support or a cache, loads into RAM, and scores audio on the CPU, GPU, or a neural accelerator. Quality depends on which model you picked and whether your machine can run it in real time. Smaller models dictate faster and miss more; larger models are slower and hungrier. If the model never loaded, there is nothing to run offline — the download is a one-time online step, not a permanent tether.
People want offline mode for flights, trains, rural links, and for a privacy posture where dictation audio never traverses a network. It is not automatic anonymity: the text still exists on disk if the app keeps history, and the focused app may sync the resulting document elsewhere. Offline also does not mean you lack a cloud option in the same product. Some tools default to local recognition and offer a network engine as a switch for older hardware.