Qwen is likely not going to be the agent for this - is 3.5 9b the best to run things local? Alternatively, what are the best ways to get privacy-preserving inference?