Accurate Bill of Quantities (BOQ) estimation requires predicting multiple construction quantities whose relationships can differ across materials, fixtures, plumbing components, and labor. This paper proposes a target-wise adaptive multi-model ensemble for fourteen-target BOQ regression. Instead of selecting a single globally strongest model, eight heterogeneous regressors—Ridge, Random Forest, MLP, XGBoost, LightGBM, CatBoost, FT-Transformer, and Cross-Target Dependency Network (CTDN)—are trained and combined through target-specific non-negative weights constrained to sum to one. The weights are learned from a validation subset using inverse-validation-MSE initialization followed by constrained Sequential Least SQuares Programming (SLSQP) optimization, while a separate validation subset is reserved for conformal calibration and the test set remains untouched until final evaluation. On the held-out test set, the proposed ensemble achieves an average R² of 0.9190, compared with 0.9188 for CTDN, 0.9188 for the existing XGBoost–FT-Transformer weighted ensemble, and 0.9185 for simple averaging. The learned weight matrix remains distributed across multiple models, with an average of 4.36 active models per target, a mean maximum weight of 0.4345, and mean normalized entropy of 0.6592, supporting target-specific model complementarity rather than single-model selection. Split conformal prediction applied to the adaptive ensemble achieves 94.83% empirical coverage at a 95% nominal level. The results support target-wise adaptive aggregation as a methodological approach for multi-target BOQ estimation while providing calibrated prediction intervals.