It use openai/privacy-filter which is smaller than 1GB in size. I haven't checked usage during inference. It rans at 1k toks/sec on my macbook. I will update the repo with benchmark results. Thanks for the comment