In deep learning (DL) development, faults may originate from training program bugs or issues in the training data. Prior work shows that simulating such faults using pre-training mutation operators can produce realistic mutants. However, this realism comes at high computational cost due to repeated retraining. Post-training mutation avoids retraining, yet relies on a fixed set of predefined operators that struggle to capture the diversity of real faults and are sensitive to training stochasticity. To address these challenges, we propose a new direction for DL mutation testing by framing post-training mutation as a model-to-model transformation problem. Instead of manually designing mutation operators, we treat mutation generation as a learned transformation from an original trained model to its corresponding faulty or realistically mutated trained model. Our approach fine-tunes a pre-trained code model to learn this transformation, enabling direct manipulation of trained model parameters without retraining. As a proof of concept, we fine-tune CodeT5+ on a small neural network using controlled original–mutated model pairs. Even in a few-shot fine-tuning setting, the model successfully generates valid mutated models by modifying the parameters of the original trained model. These results demonstrate the feasibility of learning model-level mutations through model-to-model transformations. However, sequence-based representations face scalability limitations for larger networks. This motivates a transition toward graph neural networks that better align with the structural representation of deep neural models. This work lays the foundation for scalable, efficient, and realistic mutation testing for DL systems.