They have the option to use a local llm but default to online inference. Local is not good enough for most people and people would be really unhappy having their browser use 10gb of ram and spin their fans every query.

Minstrel is working on specialist small models. Maybe one day they are small enough to run in the background no worries.