the ner.teach recipe will update as you are annotating with the purpose of finding more high quality examples to annotate first. That means that if you provide it a base model that it will be finetuned in the loop. However, once you're done annotating it won't save the updated model to disk. If you want a finetuned model on disk you'll probably want to use the train command after annotating some examples. This will train a model on your new data, saved on disk so that you may re-use it elsewhere.
Does this help? Feel free to ask more questions if not.