Nous Research said Arete-27B recorded a 1% result on its propaganda-rate measure after training, compared with 22% before, a comparison the company cited in describing efforts to shape the model’s political responses.
Arete-27B is built on a model from Alibaba’s Qwen family. Nous Research said it continued training that existing system after its initial development, making the project a post-training effort. Post-training is additional work applied after a model’s original training.
Qwen is an open-weight language-model family from Alibaba, and its availability allows developers to adapt models in the series for further work. Arete-27B extends an existing foundation through additional training.
Nous Research said the project was intended to make the model less aligned with narratives it associates with the Chinese state. The company pointed to Western liberal norms as a reference for the responses it wanted the system to produce. The percentages are Nous Research’s reported readings.
