The paper introduces MarS, a financial market simulation engine powered by the Large Market Model (LMM), a generative foundation model designed for order-level market simulation. LMM combines two complementary approaches: Order Sequence Modeling using a causal transformer to capture individual order sequences, and Order-Batch Sequence Modeling using an auto-regressive transformer with VQ-VAE to model aggregated market behaviors. The model employs a novel tokenization strategy that incorporates order attributes and Limit Order Book (LOB) information, enabling high-resolution simulation of market microstructures.

MarS leverages LMM to generate realistic market trajectories through a conditional generation process that considers four key conditions: scenario descriptions (DES_TEXT), user-injected interactive orders, starting sequences, and market matching rules (MTCH_R). The system includes a simulated clearing house that matches orders in real-time, creating a dynamic feedback loop for continuous simulation. Experiments demonstrate that MarS successfully replicates 14 stylized facts of financial markets and shows strong scalability, with performance improving as model size increases. The paper showcases MarS's applications in four domains: forecasting (outperforming DeepLOB baseline), detection (identifying market manipulation patterns), 'what-if' analysis (evaluating market impact), and reinforcement learning agent training (providing realistic environments for strategy development). This represents a paradigm shift in financial simulation, moving from traditional statistical approaches to generative foundation models that can handle complex, interactive market scenarios.