Nous Research cuts propaganda in a Qwen model from 22 to 1 percent

Nous Research, an AI lab in Austin, Texas, changed Qwen3.8-27B, an open-weight model from the Chinese company Alibaba. In an open-weight model, the weights are open to all. The new model is Arete. Nous tells that Chinese models hold the views of the Chinese government in their weights. Nous tells that the change cut answers with propaganda from 22 percent to 1 percent.

How it works

Nous gave the model example answers to learn from. It used 50 million tokens, which are small pieces of text. The teacher model was Inkling, an open-weight model from Thinking Machines Lab. Nous tells that the propaganda layer is small and easy to remove.

The numbers

Nous tells that its fact score went from 2.32 to 2.98 out of 5. Nous made the test. Hirundo makes data removal tools. It tells that it cut such output from 89.8 percent to 2.8 percent in a different Qwen model. A test by Booz Allen Hamilton found more code with errors in some Chinese models.

What is not known

The term “Western value” is not clear. Nous did not get the large drop that Hirundo reports. No external group tested Arete. A reporter tested a Qwen variant with a question about Tiananmen Square in 1989. The model did not answer, or gave a weak answer.

Insight: Western labs must start from Chinese models

Nous tells that most good open models are Chinese, and only open models let other labs change them. It found views of the same type in the Nemotron 3 Ultra model from Nvidia. Alibaba also puts Qwen in Apple Intelligence in China from December. This reads as Chinese weights as the first step of open work, and each repair must use them. Find who made the first version of each model you use.