A Chinese model retrained on 'Western values': Nous Research's Arete-27B cuts its propaganda rate from 22% to 1%

"The values in an open model are one fine-tune away." Nous Research's Nov. 7 announcement passed about 2,200 likes and 64,000 views in under a day. Arete-27B is Alibaba's Qwen3.8-27B post-trained to reinforce Western values.

According to Nous Research, Chinese open models are among the strongest available but ship with state-aligned behavior in the weights. The team built an 803-prompt eval across seven value categories, including freedom of expression, democracy and privacy, and measured how often a model adopts the Chinese Communist Party line. After fine-tuning on roughly 53 million tokens, the "propaganda rate" fell from 22% to 1%, in line with the Western models tested, with no significant loss on Hermes Bench. The eval, weights and training data are released.

The blog cites US restrictions on federal use of Chinese models such as DeepSeek, and also notes that frontier labs share a similar worldview and have shown political bias in the past. When a user wrote that 53 million tokens to change a model's politics is far fewer than expected, Nous replied that its research suggests the bias layers in Chinese-aligned models are shallow and applied late in training. The figures come from Nous's own eval and have not been independently verified.