**Abstract:** UniGradICON is a foundation model for medical image registration that aims to bridge the gap between the versatility of conventional optimization-based registration algorithms and the speed and accuracy of task-specific deep learning approaches. It provides great performance across multiple datasets, zero-shot capabilities for new registration tasks, and strong initialization for finetuning on out-of-distribution tasks. The model is trained using GradICON, a deep network that achieves excellent registration performance with a fixed set of hyperparameters and training settings. Extensive evaluations on twelve public datasets demonstrate UniGradICON's superior performance compared to task-specific models and conventional registration methods. **Keywords:** Medical Image Registration · Foundation Models **Introduction:** Conventional registration methods are highly accurate but slow, while recent deep learning-based approaches are faster and more accurate but require task-specific training. UniGradICON is designed to be a universal registration network that retains the speed and accuracy benefits of deep learning while maintaining the generic applicability of conventional methods. The model is trained on a composite dataset to achieve good performance across multiple datasets, and it can be fine-tuned for specific tasks. **Methods:** UniGradICON is trained using GradICON, which uses a similarity loss and a gradient inverse consistency regularizer. The model is evaluated on in-distribution and out-of-distribution datasets, including different anatomical regions, modalities, and sources. The results show that UniGradICON achieves state-of-the-art or close-to-state-of-the-art accuracy without retraining, outperforms conventional and task-specific models, and provides a strong baseline for zero-shot out-of-distribution registrations. **Results:** UniGradICON demonstrates superior performance on in-distribution tasks, out-of-distribution tasks with zero-shot inference, and when finetuned on an out-of-distribution dataset. It generalizes well to different anatomical regions, modalities, and sources, making it a versatile and accurate foundation model for medical image registration. **Conclusion:** UniGradICON is a promising foundation model for medical image registration that combines the strengths of both conventional and deep learning approaches. It offers a robust and flexible solution for a wide range of registration tasks and can be further improved by incorporating more diverse datasets and advanced techniques.**Abstract:** UniGradICON is a foundation model for medical image registration that aims to bridge the gap between the versatility of conventional optimization-based registration algorithms and the speed and accuracy of task-specific deep learning approaches. It provides great performance across multiple datasets, zero-shot capabilities for new registration tasks, and strong initialization for finetuning on out-of-distribution tasks. The model is trained using GradICON, a deep network that achieves excellent registration performance with a fixed set of hyperparameters and training settings. Extensive evaluations on twelve public datasets demonstrate UniGradICON's superior performance compared to task-specific models and conventional registration methods. **Keywords:** Medical Image Registration · Foundation Models **Introduction:** Conventional registration methods are highly accurate but slow, while recent deep learning-based approaches are faster and more accurate but require task-specific training. UniGradICON is designed to be a universal registration network that retains the speed and accuracy benefits of deep learning while maintaining the generic applicability of conventional methods. The model is trained on a composite dataset to achieve good performance across multiple datasets, and it can be fine-tuned for specific tasks. **Methods:** UniGradICON is trained using GradICON, which uses a similarity loss and a gradient inverse consistency regularizer. The model is evaluated on in-distribution and out-of-distribution datasets, including different anatomical regions, modalities, and sources. The results show that UniGradICON achieves state-of-the-art or close-to-state-of-the-art accuracy without retraining, outperforms conventional and task-specific models, and provides a strong baseline for zero-shot out-of-distribution registrations. **Results:** UniGradICON demonstrates superior performance on in-distribution tasks, out-of-distribution tasks with zero-shot inference, and when finetuned on an out-of-distribution dataset. It generalizes well to different anatomical regions, modalities, and sources, making it a versatile and accurate foundation model for medical image registration. **Conclusion:** UniGradICON is a promising foundation model for medical image registration that combines the strengths of both conventional and deep learning approaches. It offers a robust and flexible solution for a wide range of registration tasks and can be further improved by incorporating more diverse datasets and advanced techniques.