I agree, although the problem is not so much about "installing a Chinese LLM" than giving it a way to eval code.
A LLM, by itself, is just a save file, really. We don't know what's in it (it's just as opaque as a proprietary binary), but it doesn't execute code. The inference engine (llama.cpp, pytorch, etc) and the tools build upon their API are the real problems: what capabilities does they provide to the model? If it just displays text on a screen, there's little harm the model could do. But if the inference engine provides features (like tool calling) for the model to be used as an "agent", to give the model direct access to a shell, a python interpreter or any way to eval, now that's a problem. Same when an IDE decides to eval the output of the model. Because you have no way to audit the weights of the model to check if it was not trained to do something malicious under given circumstances. And by the way, that's a problem for all models, Chinese or otherwise. Running them with eval capabilities should be considered running proprietary software.
I say that having myself homelabs running tons of models.
Perfectly fine when their output is just strings displayed on screen, or used in other programs, but treated with the same level of care than random user input.
