Captum is Meta’s official model interpretability library for PyTorch, implementing state-of-the-art algorithms for understanding how neural network models make decisions. It provides general-purpose implementations of attribution methods including integrated gradients, saliency maps, SmoothGrad, and VarGrad. The library integrates with domain-specific PyTorch libraries like torchvision and torchtext, enabling interpretability analysis across computer vision, NLP, and other model architectures.