Qwen 3.6's arrival makes the standard capability-versus-usability split in open-source AI concrete in a way that abstract arguments have not managed. The model's architecture — activating only 3 billion of its 35 billion parameters per query — is engineered for efficiency on consumer hardware, and the . But the practitioners who deployed it immediately encountered a different constraint: a model that loops on instruction confirmation, substituting its own judgment for the user's stated preference in a cycle with no natural exit .