Thank you very much for your response and detailed explanation of possible options. I will try the same. Your point about dynamic shapes was very insightful as I was thinking the opposite. I am already exploring layer wise execution as an alternative. I had a couple of queries: Are there official Core ML tools for QAT, model compression and knowledge distillation? Or is this to be done before convering to an .mlprogram? If no, what would be the workflow (in terms of libraries and tools) you would suggest to perform these operations so that the resultant model is compatible with Core ML conversion