Structurally cheaper inference for open models.