We introduce Trankit, a light-weight Transformer-based Toolkit for\nmultilingual Natural Language Processing (NLP). It provides a trainable\npipeline for fundamental NLP tasks over 100 languages, and 90 pretrained\npipelines for 56 languages. Built on a state-of-the-art pretrained language\nmodel, Trankit significantly outperforms prior multilingual NLP pipelines over\nsentence segmentation, part-of-speech tagging, morphological feature tagging,\nand dependency parsing while maintaining competitive performance for\ntokenization, multi-word token expansion, and lemmatization over 90 Universal\nDependencies treebanks. Despite the use of a large pretrained transformer, our\ntoolkit is still efficient in memory usage and speed. This is achieved by our\nnovel plug-and-play mechanism with Adapters where a multilingual pretrained\ntransformer is shared across pipelines for different languages. Our toolkit\nalong with pretrained models and code are publicly available at:\nhttps://github.com/nlp-uoregon/trankit. A demo website for our toolkit is also\navailable at: http://nlp.uoregon.edu/trankit. Finally, we create a demo video\nfor Trankit at: https://youtu.be/q0KGP3zGjGc.\n