Researchers from Shanghai AI Laboratory have introduced InternW0, a foundational physical world model designed for efficient, actionable real-world interactions. The architecture utilizes an asymmetric video-action framework driven by flow matching, which integrates a high-capacity video expert for long-horizon predictive context with a lightweight action expert operating at a faster timescale. To ensure computational efficiency, the model employs observation-conditioned context routing, which reuses layerwise Key/Value (K/V) states to adapt to new observations instead of regenerating the entire future sequence for every action update. InternW0 was trained on approximately 7,200 hours of heterogeneous robot and egocentric data, including the 275-hour EgoLab dataset. The model supports heterogeneous embodiments through domain-specific interfaces and soft prompts, while contact-aware post-training enables the integration of force and tactile signals for complex manipulation tasks. Empirical evaluations demonstrate the model's efficacy across simulation benchmarks and sophisticated real-world scientific workflows, such as a 15-stage metal-organic framework synthesis and a 5-stage dexterous manipulation task for quantitative pipetting. By combining asynchronous multi-frequency processing with local physical modeling under partial observations, InternW0 establishes a scalable framework for science-native, universal physical intelligence.