Use mask rcnn models in production

[rusanager]

Hello!
I have several mask rcnn models whose trained with enough large datasets including high-resolution images. My models are large, about 170-270MB.
I would like to know how I can optimize my model and the best practices for using large models in production. I need to do several recognitions step by step and it takes about 1 minute. I need to scale my server however I don't get how I can create new instances quickly and save the speed of recognition. If I will save models outside the server I need to download about 1GB for recognition.
I will be glad to hear any piece of advice.