Comment by cubefox

"Smaller" models still require excessive RAM while not producing acceptable results for conversation. LLMs will probably always run in the cloud, anything else will just be too weak for most applications.