Arete-27B: Post-Training Qwen to Reduce State-Aligned Bias

Nous Research has post-trained Alibaba's Qwen3.8-27B model to create Arete-27B, aiming to reduce Chinese state-aligned bias and reinforce Western values. On the organization's evaluation of 803 prompts across seven value categories, the new model raised its fact score and lowered its propaganda rate while retaining its core agentic and general capabilities. The authors argue that the bias layer in Chinese-aligned models is shallow and can be removed through careful supervision without disturbing underlying capabilities. Open article →