API
< Back to all clusters
[TECHNOLOGY] · United States, China · 3 sources

started · updated

Nous Research releases Arete-27B to mitigate Chinese AI model bias

Nous Research has released Arete-27B, an open-weight model based on Alibaba Cloud’s Qwen3.8-27B, designed to reinforce Western values and reduce political bias. The developer stated that Chinese AI models may contain biases in political topics or output vulnerable code when used by US organizations.

To address this, Nous Research applied post-training using a dataset of approximately 52.8 million tokens across seven categories, including freedom of expression and national ideology. Evaluations conducted using SpaceXAI’s Grok 4.7 showed that Arete-27B increased factuality scores from 2.32 to 2.98 compared to the base Qwen model, while reducing propaganda content from approximately 22% to 1%.

While the release has been praised by some, it has also faced criticism for potentially embedding the personal political views of developers under the guise of Western values. Nous Research responded by stating the project demonstrates that anyone can tune the values of an open model.

Entities

Alibaba Cloud · Arete-27B · Nous Research · Qwen3.8-27B · SpaceXAI