The advancement of Embodied AI heavily relies on large-scale, interactive 3D scene datasets characterized by scene diversity and realistic layouts. However, existing datasets typically suffer from limitations in diversity or simulatability, sanitized layouts lacking small items, and severe object collisions. To address these shortcomings, we introduce InternScenes, a novel large-scale interactive indoor scene dataset comprising approximately 40,000 diverse scenes by integrating three disparate scene sources — real-world scans, procedurally generated scenes, and designer-created scenes. It includes 1.96M objects and 800k CAD models that cover 15 common scene types and 288 object classes, resulting in complex layouts with an average of 41.5 objects per region — the highest to date. Our comprehensive data processing pipeline ensures simulatability by creating real-to-sim replicas for real-world scans, achieves realistic layouts by preserving small items, and enhances interactivity by incorporating interactive objects and resolving collisions. We demonstrate the value of InternScenes through two benchmark applications: scene layout generation and point-goal navigation. Both highlight the new challenges posed by the complex and realistic layouts. More importantly, InternScenes paves the way for scaling up model training for both tasks, enabling generation and navigation in such complex environments.