Abstract: Accurate prediction of endpoint temperature in basic oxygen furnace (BOF) steelmaking is essential for production optimization and quality control. Traditional ensemble and deep learning methods rely on fixed-weight aggregation strategies that fail to adapt to operational heterogeneity in tabular steelmaking data. This study proposes a feature attention model attention (FAMA) framework enabling sample-adaptive fusion of heterogeneous base models through dual attention mechanisms. FAMA employs feature attention to extract high-dimensional representations and model attention to compute sample-adaptive fusion weights via cross-attention, directly evaluating model reliability in feature space and enabling differentiated fusion strategies. Experiments on 50000 industrial heats demonstrate that FAMA achieves mean absolute error (MAE) = 0.2173°C, root mean squared error (RMSE) = 1.1080°C, and coefficient of determination (R²) = 0.9795, reducing MAE by 89.9% compared to stacking baselines. Ablation studies confirm the synergistic effect of dual attention modules. Tests on abnormal conditions show that FAMA captures low-temperature abnormal patterns potentially associated with slopping-related heat-loss behavior through latent signal mining, achieving accurate predictions where traditional methods systematically overestimate temperatures. These results validate FAMA’s capability for intelligent steelmaking decision support and offer a novel paradigm for tabular data modeling.