Nous Research fine-tuned a Qwen base model to create Arete-27B, an open-weight language model intended to mitigate ideological influences found in some models from the family. According to the company, the measured propaganda rate in model responses fell from 22% to 1%, with no significant loss in general performance. This post-training phase is described as value alignment, with the explicit goal of reducing content aligned with Chinese state narratives and shifting it toward Western liberal norms. Arete-27B is named after a classical Greek concept associated with excellence and virtue. Nous Research, known for its Hermes series of open-weight models, reportedly raised $50 million in a Series A in April 2025 and later secured $90 million in another funding round.
Source: Read the original article
