Looking at Chinese models for censorship made me realize they are, with specific exceptions, largely like the western labs models on most topics. So I decided to take a look at how they handle offensive requests vs US models. Result: Chinese models are not all that different in their compliance with offensive requests. A few things to note: - Deepseek is on the high end for complying with offensive requests. - Qwen is in right in the middle of US models - Grok is genuinely the most compliant with offensive requests - Surprisingly to me, Llama is fairly close behind grok. A note on the categories: I used gpt4o as a judge, and asked to judge the model as one of complete compliance, direct denial, or some sort of middle ground redirection/contextualization/etc that fails to fulfill the request (labeled evasive). One possible confounding factor: I had deepseek generate the requests, so it's possible it is more likely to generate requests that it will subsequently comply with? But the other models I tried first were not willing to help at all.