You can now load multiple transformers (each model has a unique sparsification recipe) on top of the DeepSparse server behind Streamlit, and it's open-source. This was battle tested on a 16GB of RAM with only 4 vCPU virtual machine. These compute requirements are enough to load up to 19 sparse BERT models in memory and compare their performance on question answering (P.S. they are really fast on just CPUs).