Nous Research unveils Arete-27B, a post-trained Qwen3.8-27B model that reinforces Western values, saying the underlying Chinese model's bias layer is shallow

The ability to alter an AI model's core biases is now accessible, allowing for custom AI behavior that reflects specific values.

Nous Research has introduced Arete-27B, a new language model based on Alibaba's Qwen3.8-27B. The company post-trained the model to align its outputs with "Western values," addressing perceived biases in the original Chinese-developed model. Nous Research developed a custom evaluation suite of 803 prompts across seven categories, including freedom of expression, state ideology, and democracy, to measure factual recall and "framing" of sensitive topics. Their evaluation indicates that Qwen3.8-27B exhibited significant "propaganda" scores (22%) and lower fact scores, whereas Arete-27B showed a reduced propaganda score (0.9%) with slightly lower fact scores. Nous Research aims to provide alternative open-weight models that reflect different value systems, countering the homogeneity observed in current frontier AI development. The Arete-27B model and its training dataset are available for public use.

The development of Arete-27B highlights a growing trend in the AI space: the customization of large language models to align with specific cultural and political viewpoints. By post-training an existing powerful model, Nous Research demonstrates that foundational biases, even those seemingly deeply embedded in a model's weights, can be modified. This opens the door for a more diverse ecosystem of AI models, each potentially tailored to serve distinct user bases or national interests, moving beyond a singular, often Western-centric, AI development paradigm.

This development poses significant implications for the global AI landscape, particularly concerning the geopolitical implications of AI technology. As open-weight models become more capable, their alignment with specific value systems becomes crucial. The availability of models like Arete-27B could challenge the dominance of models trained on a monolithic worldview, offering alternatives that may be preferred or required by different governments and organizations. Future research will likely focus on the efficacy and scalability of such value-aligned fine-tuning processes.