inference speed of the models is probably the bottleneck