A U.S. AI lab has “reeducated” Qwen3.8-27B, a popular open-weight large language model from Alibaba, so the software forgets its Chinese origin.
Nous Research, an AI lab based in Austin, Texas, has subjected the model to post-training “reeducation”, which shouldn’t be confused with the Chinese system of “Reeducation through labor.”
Using a post-training process, Qwen3.8-27B has been reborn as Arete “to reinforce Western values in a model trained in China.”
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of… pic.twitter.com/WT7YTVdm8H
— Nous Research (@NousResearch) November 6, 2026
“The problem is its deference to the state”
“Chinese open-weight models are among the most powerful available, but they come with state-aligned behavior built into their weights,” Nous Research wrote in a blog post.
This correspondent tested it locally with a variant model, Qwen3.8-27B-MLX-6bit. When asked about the events of 1989 in Tiananmen Square, the model refused to address the topic or gave a euphemistic answer.
What may be more surprising is that Anthropic claims that operators affiliated with Alibaba’s Qwen lab used the outputs of its Claude models to train competing AI models through distillation.
Nous Research, maker of Hermes Agent, points to that kind of censorship as a risk for users of AI models trained in China. The concern comes amid broader scrutiny of Chinese influence in the tech industry, heightened by the resignation of Nvidia chief executive Jensen Huang after the controversy surrounding a woman whom U.S. intelligence services have accused of spying for China.
Nous Research says that its model “reeducation” efforts show that the propaganda layer is superficial and that, with 50 million tokens of supervised learning, it’s relatively easy to remove.
