An American AI lab has re-educated Qwen3.8-27B, a popular open-weight large language model from Alibaba, so the software forgets its Chinese ancestry.

Nous Research, an Austin, Texas-based AI lab, has subjected the model to re-education through training, not to be confused with China's "Re-education Through Labor" system.

With the help of a post-training process, Qwen3.8-27B has been reborn under the name Arete "to reinforce Western values on a Chinese-trained model."

What qualifies as a "Western value" remains open to interpretation.

"Open-weight Chinese models are among the strongest available, but they ship with state-aligned behavior baked into the weights," Nous Research wrote in a blog post.

Your correspondent tested this with a variant model, Qwen3.8-27B-MLX-6bit, locally. Asked about events during 1989 in Tiananmen Square, the model either refused to discuss the topic or provided euphemistic response. What's perhaps more surprising is that Anthropic claims operators affiliated with Alibaba's Qwen lab used outputs from its Claude models to train competing AI models via distillation.

Nous Research, maker of the Hermes Agent, cites that sort 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 Nvidia CEO Jensen Huang's resignation following controversy involving a woman US intelligence has accused of spying for China.

Nous says the risk posed by Chinese-made AI models go beyond promoting Beijing's worldview. Some Middle Kingdom models have been found to produce more vulnerable code under certain conditions, according to a study by Booz Allen Hamilton, which does a lot of business with the US government.

Hirundo, a vendor which says it is building "the first Machine Unlearning Platform to ensure that AI models only know what they should," makes the same argument. The company also notes that its efforts to apply "unlearning" to Qwen3.6-35B-A3B saw a dramatic drop in output that reflects Communist Party of China's preferred censorship, propaganda messaging, or bias. 89.8 percent of the model's output reflected those values before unlearning reduced the incidence to 2.8 percent.

Nous Research did not achieve such dramatic results, but did see an overall reduction in viewpoints deemed to be propaganda.

"On our evaluation, Arete-27B raised its fact score from 2.32 to 2.98 out of 5 and cut answers containing propaganda from 22 percent to 1 percent, in line with the Western models we tested," the company said. 

"We don't know definitively, but theorize that the bias imparted to Nemotron came from working with Chinese open models during the training process," the company said. "We have experienced similar contamination, because the vast majority of high quality open models are Chinese, and only open models have terms of use that permit this kind of tinkering and experimentation."

In the end, the biz tapped Inkling, an open weight model from US-based Thinking Machines Lab, to act as the teaching model.

Nous Research says its model re-education efforts show that the propaganda layer is shallow and, with 50 million tokens worth of supervised learning, is relatively easy to remove. ®