Also, running models locally requires good hardware to get acceptable performance. It's still a large barrier to entry.