Local models look after themselves
Local models now stay within your Mac's memory, unload when you stop using them, restart after a crash, and run only once Bandicot has checked them.
Within memory
When a new model needs room, Bandicot unloads the one you used least recently, never one that is answering. Models you have not used for 15 minutes unload on their own, and one that crashes is started again.
Less memory, same answers
Three chat models now store what they have read at half the size, which frees about 2 GB on an 8B model with almost no change in its answers.
Only what Bandicot checked
Every model server Bandicot starts is first checked against the build Bandicot ships, and its chat and text servers answer only Bandicot.