You point it at a recording or a microphone and it types out what was said, in English, Chinese, French, Italian, Korean, Portuguese, or Vietnamese. The whole model fits in about 1.6 gigabytes and keeps up with live audio using three or four threads of a normal laptop processor, no graphics card involved. Nothing leaves your machine, which is the point: the audio of a client call or a doctor's dictation never touches a server.
The trick is aggressive compression. Microsoft stored most of the model's weights at close to two bits each, a technique called BitNet, then wrote custom code so ordinary processors can chew through those tiny weights quickly. The result runs 1.6 to 2.3 times faster than Whisper.cpp, the current default for local transcription, and on messy meeting recordings it also makes fewer mistakes than Whisper.