On some days, half of my promts to Copilot at work are spoken to a local whisper medium model running on an Intel ARC GPU.