This paper introduces HiSim, a hybrid framework for simulating social media user behavior in large-scale social movements. The framework categorizes users into two types: core users, driven by Large Language Models (LLMs), and ordinary users, modeled using agent-based models (ABMs). Core users represent influential individuals, while ordinary users are modeled to simulate large-scale interactions efficiently. A Twitter-like environment is constructed to replicate user response dynamics following trigger events. A benchmark called SoMoSiMu-Bench is developed to evaluate the simulation, including three real-world datasets (Metoo, RoeOverturned, and BlackLivesMatter) and evaluation strategies at both micro and macro levels. The framework addresses the challenges of simulating large-scale social movements by combining LLMs for core users and ABMs for ordinary users. This hybrid approach enables efficient and effective simulation of user behavior, capturing both individual and collective attitudes. The simulation environment includes a message feed mechanism and offline news feed to replicate real-world social movement scenarios. The framework is evaluated using micro alignment metrics (stance, content, and behavior alignment) and macro system metrics (static attitude distribution and time series of average attitude). Experiments show that the hybrid framework outperforms pure ABMs in terms of both static and time series measures. LLM-empowered agents effectively model core users' stances and content generation, while ABMs handle ordinary users' behavior. The framework is scalable, with performance and runtime variations observed across different numbers of agents. The simulation also demonstrates the ability to replicate echo chambers and test strategies to mitigate polarization. The paper concludes that the hybrid framework provides an effective and flexible approach for simulating large-scale social movements, offering insights into user behavior and collective attitudes. The framework is supported by extensive experiments and benchmarking, demonstrating its effectiveness in capturing real-world social dynamics.This paper introduces HiSim, a hybrid framework for simulating social media user behavior in large-scale social movements. The framework categorizes users into two types: core users, driven by Large Language Models (LLMs), and ordinary users, modeled using agent-based models (ABMs). Core users represent influential individuals, while ordinary users are modeled to simulate large-scale interactions efficiently. A Twitter-like environment is constructed to replicate user response dynamics following trigger events. A benchmark called SoMoSiMu-Bench is developed to evaluate the simulation, including three real-world datasets (Metoo, RoeOverturned, and BlackLivesMatter) and evaluation strategies at both micro and macro levels. The framework addresses the challenges of simulating large-scale social movements by combining LLMs for core users and ABMs for ordinary users. This hybrid approach enables efficient and effective simulation of user behavior, capturing both individual and collective attitudes. The simulation environment includes a message feed mechanism and offline news feed to replicate real-world social movement scenarios. The framework is evaluated using micro alignment metrics (stance, content, and behavior alignment) and macro system metrics (static attitude distribution and time series of average attitude). Experiments show that the hybrid framework outperforms pure ABMs in terms of both static and time series measures. LLM-empowered agents effectively model core users' stances and content generation, while ABMs handle ordinary users' behavior. The framework is scalable, with performance and runtime variations observed across different numbers of agents. The simulation also demonstrates the ability to replicate echo chambers and test strategies to mitigate polarization. The paper concludes that the hybrid framework provides an effective and flexible approach for simulating large-scale social movements, offering insights into user behavior and collective attitudes. The framework is supported by extensive experiments and benchmarking, demonstrating its effectiveness in capturing real-world social dynamics.