FlashPack is an innovative file format and loading mechanism designed to significantly speed up model checkpoint I/O in PyTorch, offering 3-6× faster loading times compared to existing methods like accelerate and standard load_state_dict(). By flattening a model's state_dict into a single data stream and utilizing memory-mapped reads with overlapping disk, CPU, and GPU processes, FlashPack eliminates the synchronization delays and overhead typical in current model loading processes. This pure-Python package is compatible with systems lacking GPU Direct Storage and works by reconstructing tensors directly in GPU memory without data copying. Despite its impressive performance, FlashPack has limitations, such as requiring weights of the same data type and lacking support for pipeline parallelism or state dictionary transformations. It can be easily integrated into existing workflows through mixins or direct calls and is accessible via PyPI or GitHub.