They definitely do. A common configuration is running a supervisor model in the cloud and a much smaller model locally to churn on long running tasks. This frees Openclaw up to lavishly iterate on tool building without running through too many tokens.