This version is a major architectural upgrade to V2.X - the neural POS tagger, morphological tagger and sentence splitter have been replaced with a single transformer based model (work by nitinvwaran) trained on all three tasks using Multitask Learning (MTL). This lowers the memory footprint of the system, substantially speeds up start up and run-time, and allows the system to run on GPU with a more modest GPU RAM (~4 GB). Accuracy is on par with or slightly better than the previous version.