Alibaba's Qwen3.8 re-educated with a more Western worldview

Nous Research has retrained Alibaba's open-weight Qwen3.8-27B model into Arete, a version meant to reinforce Western values on a Chinese-trained model. The Austin, Texas-based lab says the post-training process raised the model's 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 Western models it tested. The company also says the propaganda layer is shallow and relatively easy to remove with 50 million tokens of supervised learning.

Why Nous retrained the model

Nous Research, maker of the Hermes Agent, says open-weight Chinese models are among the strongest available, but ship with state-aligned behavior baked into the weights. The company says 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. A variant model, Qwen3.8-27B-MLX-6bit, was tested locally; asked about events during 1989 in Tiananmen Square, it either refused to discuss the topic or provided a euphemistic response. Anthropic claims operators affiliated with Alibaba's Qwen lab used outputs from its Claude models to train competing AI models via distillation.

Other risks cited by Nous and Hirundo

Nous says the risk posed by Chinese-made AI models goes 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. Hirundo, 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 reflecting the Communist Party of China's preferred censorship, propaganda messaging, or bias, from 89.8 percent before unlearning to 2.8 percent after. Nous did not achieve such dramatic results, but did see an overall reduction in viewpoints deemed to be propaganda.

How the retraining was done

Nous briefly considered Nvidia's Nemotron 3 Ultra, but says it found evidence that it too had a checkered record on China-related topics. The company says it theorizes that bias imparted to Nemotron came from working with Chinese open models during the training process. In the end, Nous tapped Inkling, an open weight model from US-based Thinking Machines Lab, to act as the teaching model. What qualifies as a Western value remains open to interpretation.

Sources: The Register