[Western AI circles are obsessed with political correctness: Instilling "Western values" into Alibaba's Qwen] Beating AI News reports that Nous Research, the developer of Hermes Agent, released Arete-27B, directly using Alibaba's Qwen3.8-27B as a foundation, specifically for training with "Western values." Nous itself admits that Chinese open-source models are currently very powerful, but their value orientations do not meet their standards. Therefore, they used approximately 53 million tokens for fine-tuning, focusing on issues such as freedom of speech, democracy, and privacy. To test the effectiveness of the modifications, Nous also created 803 political value test questions, inviting xAI's Grok 4.7 to act as judge. According to their own evaluation standards, the model's "propaganda rate" dropped from 22% to approximately 1%, while its general capabilities did not significantly decline. [Original Link]