Nous Research Says Arete-27B Cut Qwen’s Propaganda Rate to 1%

Nous Research has released Arete-27B, a post-trained language model built on a base from Alibaba’s Qwen family. The company put its propaganda-rate measure at 1% after training, compared with 22% before the work.

Nous Research described the goal as shifting responses away from narratives it associates with the Chinese state and toward Western liberal norms. The 22%-to-1% change is the company’s reported measurement.

Qwen is an open-weight family of language models from Alibaba that developers commonly use as a starting point for further work. Its open-weight availability lets developers build on the models. Arete-27B adds training to an existing Qwen model rather than introducing a new base model.

Post-training refers to steps applied after a model’s initial training, including supervised fine-tuning and reinforcement learning. The name Arete draws on a Greek ideal associated with excellence and virtue.