Propaganda-style response 22%→1% self-evaluation… Areté-27B released

Nous Research released Arete-27B, an AI model post-trained on Alibaba's Qwen, reporting that its own evaluation showed propaganda-style responses to political questions fell from 22% in the base model to 1% in the tuned model. The model was not retrained from scratch but built on Qwen's open weights with additional post-training to adjust response tendencies. Crypto Briefing reported the release at 6:53 a.m. KST on the 7th. The result matters because it suggests lightweight post-training can sharply reduce political bias in open-weight models without full retraining. Next, watch for independent benchmarks and adoption of the Qwen-based model by developers.