Ludwig 0.5 represents a significant overhaul of the open-source deep learning framework, shifting its backend from TensorFlow to PyTorch to enhance scalability, deployment, and reliability. This version introduces new features like step-based training and evaluation, data balancing, and end-to-end TorchScript export for optimized model deployment. The integration with Ray and the addition of image encoders like MLPMixer and ViTEncoder improve training speed and model performance on large datasets. AutoML capabilities for tabular and text classification have been enhanced, along with scalability optimizations using Dask, Modin, and Ray for efficient processing of large datasets. The update also includes improved configuration validation through marshmallow schemas, expanded testing for increased stability, and integration with Aim for experiment tracking. Despite extensive changes, backward compatibility has been prioritized, maintaining consistency with previous versions, though with some minor adjustments to parameter naming conventions. The update culminates in an article on the PyTorch blog, underscoring the benefits of PyTorch's pythonic design and developer-friendly ecosystem, and invites the broader deep learning community to contribute to Ludwig's ongoing development.