Models distribution, nuget or hosting? best practices?

[vpenades]

When working with ONNX models, I am faced more often than not with the dilemma of how to distribute the models related to a library.

Let's say I develop a library X that depends on a given ONNX model. Since the model is a dependency, it feels obvious to distribute the model within the nuget package of the library.

And this seems fine for small ONNX models, but as models become bigger, a number of issues appear:

• Will nuget.org accept bigger and bigger nuget packages?
• when new versions of the library are published, it makes your local cache increase size exponentially. Last time I checked my local nugets cache was over 15gb
• ONNX models can be included into nuget packages as embedded resources within the DLL or as content files, each option has pros and cons depending on the platform. For example on Android embedding a model is simply not an option due to the android app size limit.

So my question is... which would be the best practices for model distribution in libraries?