Understanding how opinions shift in response to external events is essential for modeling real-world social dynamics. While traditional agent-based models offer structural insights, they lack language understanding and struggle to adapt to open, evolving information environments. Recent advances in Large Language Models (LLMs) offer a new path forward, yet assigning an LLM agent to every user is both computationally infeasible and socially noisy. In this paper, we introduce OpinioNet, a scalable framework that models LLM-powered community agents, each representing an ideological community. This design enables efficient, context-aware simulation while maintaining social realism. To enrich agent expressiveness and mitigate over-smoothing, OpinioNet integrates multi-level personas, including abstract ideological tags, distilled group narratives, and semantically retrieved historical posts. Additionally, simulated influencer endorsements embed representative user voices into each community’s response. User-level opinions are then updated by combining external event influence, social network structure, and individual opinion inertia. Experiments show OpinioNet outperforms classical opinion dynamics models, achieving +18.29% Micro-F1, +19.21% Macro-F1 and +39.78% Pearson correlation, demonstrating a practical and interpretable solution for simulating ideological change at scale.