Abstract: A method and system are provided for training and deploying quantum foundation models to generate improved inputs for quantum computational devices. In one embodiment, training samples for a pretraining task (covering unitary matrices, quantum circuits, or microwave pulses) are automatically generated using low-complexity classical processes. A quantum foundation model is pretrained on these samples to capture device- agnostic structure associated with a family of quantum devices and the designated pretraining task. The pretrained model is then fine-tuned using a specialized dataset for a particular device and computational objective. The resulting fine-tuned model produces optimized inputs, such as circuits or microwave control pulses, that reduce gate times, gate errors, cross-talk, leakage, or gate counts when executing quantum algorithms. This approach enables large-scale model training without requiring quantum hardware during pretraining.