Directly load spacy model from packaged tar.gz file

No, that's currently not possible. The main purpose of the .tar.gz archives is to make them easy to install via pip install. However, you can always extract the model data from the archive, and then load it in from a path – see here for more details.

nlp = spacy.load('/path/to/en_core_web_md')

Using the spacy link command you can also create "shortcut links" for your models, i.e. symlinks that let you load in models using a custom name instead of the full path or package name. This is especially useful if you're working with large models and multiple environments (and don't want to install the data in each of them).

python -m spacy link /path/to/model_data cool_model

The above shortcut link would then let you load your model like this:

nlp = spacy.load('cool_model')

Alternatively, if you really need to load models from an archive, you could always write a simple wrapper for spacy.load that takes the file, extracts the contents, reads the model meta, gets the path to the data directory and then calls spacy.util.load_model_from_path on it and returns the nlp object.

Its not the direct answer but it might be helpful in order to load compressed models directly with SpaCy. This can be done by using pickle.

First, you need to load your SpaCy Model and dump it compressed with pickle:

import spacy
import pickle

s = spacy.load("en_core_web_sm", parse=False)

pickle.dump(s, open("save.p", "wb"))

Afterwards, you can load easily somewhere else the pickle dump directly as SpaCy model:

s = pickle.load(open("save.p", "rb"))