We adapted a modern self-supervised learning objective Dinov3, originally developed for images, to ECGs with an arbitrary number of leads. We opted for a Vision Transformer for our model architecture. The Dinov3 objective enabled our model to learn meaningful global and dense signal representations without relying on manual labels during training. This allowed the model to benefit from large datasets that would otherwise be too costly to annotate. To ensure robustness, we explicitly trained the model to handle common disturbances that appear in real-world recordings. These include electrode motion artifacts, baseline drifts, and muscle interference. We achieved this by augmenting the training data with realistic noise patterns taken from the MIT-BIH noise stress test database. As a result, the model learned to focus on clinically relevant signal features rather than being distracted by noise. Our training dataset combined several large and diverse collections of ECGs originating from different hospitals, devices, and recording protocols: The Harvard Emory ECG Database, the CODE database, and the Incentia11k database. Importantly, the data included recordings with varying numbers of leads, which allowed the model to generalize across different sensor configurations. Training the model required distributing both the data and its parameters across multiple CUDA-enabled GPUs. We used the PyTorch library and a fully sharded data parallel approach to efficiently handle memory usage and computation.