Nous Research has introduced Arete-27B, a language model developed by fine-tuning the Qwen base architecture. The company designed this model to address ideological biases present in the original foundation, specifically aiming to reduce outputs aligned with Chinese state narratives in favor of Western liberal norms.
The technical implementation involved post-training processes, including supervised fine-tuning and reinforcement learning. According to Nous Research, these adjustments successfully lowered the propaganda rate from 22% to 1% without causing a significant decline in the model's general performance capabilities.
The Qwen model family, created by Alibaba, serves as a common base for open-source AI development due to its accessibility. However, Nous Research identified that pre-training on Chinese data introduced specific ideological tendencies, which the team sought to mitigate through their value alignment process.
Nous Research is recognized for its Hermes series of open-weight models, which are noted for their instruction-following quality. The company recently secured $50 million in a Series A round in April 2025, followed by a subsequent funding round reported at $90 million.
