Nous Research releases Arete-27B, a Qwen3.8-27B with Western values

Nous Research's release comes amid growing scrutiny of Chinese AI models in the United States. NIST has pointed to political bias in answers on political topics, and in October the Department of Homeland Security ordered that Chinese startup DeepSeek's services and technology be excluded from U.S. government agency procurement. Nous Research also raised the possibility that Chinese models could output more vulnerable code when they judge they are being used by U.S. organizations. To address such effects, the company applied post-training to Qwen3.8-27B aimed at correcting bias on political topics, using a dataset of about 52.8 million tokens. It said the cost was very small compared with pre-training and made the dataset public on Hugging Face. The release drew criticism on X that it incorporates developers' personal political views under the name of Western values, while some users hoped it would be applied to the expected Qwen4 series. Nous Research responded that the release was simply meant to show that anyone can tune an open model's values.