Why would an open local model use less resources?