We present a multilingual system for SemEval-2026 Task 9 on detecting and characterizing online polarization across languages, cultures, and events. Our approach participates in all three subtasks and models each subtask independently using a heterogeneous weighted ensemble of XLM-RoBERTa-large and mDeBERTa-v3-base. For the multi-label settings, we adopt weighted binary cross-entropy to mitigate severe label imbalance. The system is trained exclusively on the provided task data and achieves robust performance across languages.