AI lab says a small amount of extra training stripped Chinese government talking points from Alibaba's Qwen model

Nous Research, an AI research company, released a free AI model on Nov. 6 that it says strips out much of the Chinese government's official messaging from a popular Chinese model, according to the company's announcement and its model page on Hugging Face, a site where AI models are shared.

What it changed

The new model, Arete-27B, is a retrained version of Qwen3.8-27B, made by Alibaba, the Chinese tech giant.  Qwen is an open-weight model, meaning anyone can download it, run it on their own computers and change it.  That is different from ChatGPT or Claude, which people reach only over the internet and cannot modify.  The 27B means it has 27 billion parameters, the internal settings a model learns in training; that is small enough for a powerful personal computer.

Chinese AI models are among the strongest open models, but they are built to follow Beijing's rules on sensitive topics, Nous says.  The lab says it saw Qwen refuse, dodge or give Communist Party-style answers on topics such as free speech, democracy and surveillance.

What the results show

Nous retrained the model on about 28,900 example answers written by another open AI model.  On its own test of 803 questions, answers that adopted the Chinese state's version of events fell from 22% to about 1%, similar to Western models it tested.  The biggest change came on freedom of expression, where such answers fell from 60% to 1%.  The lab says scores on its general skills tests stayed about the same.

The results come from Nous's own test, graded by an AI judge, and have not been reviewed by outside researchers.  Nous says the main lesson is that the political layer in such models is shallow, so the values in any open model can be changed with a small amount of extra training.

Source:  Nous Research announcement; Arete-27B on Hugging Face