Running large language models locally on consumer hardware has always meant accepting a strict set of trade-offs. The situation was clear: if you wanted frontier-level reasoning, you needed a cloud connection.