Its own engine
Nova runs llama.cpp as a hidden background process, only while it's open. One model is loaded at a time, so each gets the whole GPU, and idle models unload after five minutes.
Models load in a way that keeps a loaded model to about 1.7 GB of RAM instead of around 9 GB, with the rest in video memory. Closing Nova stops everything, and Windows stops the engine even if Nova crashes.







