Nous Research cuts Chinese bias in a Qwen model to 0.9 percent

Nous Research, a US startup, released Arete-27B, an open model made from Qwen3.8-27B, a model from the Chinese company Alibaba. Nous tells that training on 53 million tokens cut propaganda answers from 22 percent to 0.9 percent. The propaganda follows the Chinese government. Nous tells that the model keeps its general skills. Nous made the test.

How it works

Nous used supervised fine-tuning. This is more training on example answers. It did not use reinforcement learning. The teacher model was Inkling from Thinking Machines Lab. The weights, test data and training data are public for all to use.

The numbers

The test has 803 prompts in seven topics. Grok from xAI was the judge. The score for facts went from 2.32 to 2.98 out of 5. In free speech, propaganda went from 60 to 1 percent. On the Hermes Bench of Nous, the score went from 65.77 to 65.60 percent.

What is not known

No external group did the test again. Nous tells that Nemotron from Nvidia also showed signs of Chinese bias, but it gave no figure. Other Chinese models had 15 and 16 percent propaganda on the same test.

Insight: Model values are now a setting that a buyer can change

Nous tells that the bias in Qwen is thin. Alibaba tells that Qwen goes into Apple Intelligence for users in China from December. Vals AI found that Thomson Reuters also retrained a small Qwen model. This reads as a move of control: the team that does the last training, not the maker, sets the values. Leaders must find out who did it.