Computational engineering in industrial applications is currently constrained by the trade-off between speed and accuracy. While high-fidelity solvers provide precision, they are too computationally intensive for real-time interaction or optimization tasks. Conversely, traditional data-driven surrogate models are fast but often fail to generalize across diverse geometries or varying input parameters. To overcome this challenge, we present Large Engineering models: transformer-based reference models leveraging large amounts of simulation or experimental data. We utilize sparse attention architectures to encode both large geometries as well as parameter sets efficiently. Designed to replace numerical solvers across diverse boundary conditions, these models represent a shift towards generalized engineering intelligence. Furthermore, recognizing that domain knowledge such as governing PDEs and engineering heuristics should not be disregarded, we discuss strategies for enforcing physical consistency within such Large Engineering Models. We validate this approach through industrial showcases including injection moulding, power transformers, and hypersonic flows.