started · updated
Nous Research re-trains Alibaba Qwen model to align with Western values
The Austin, Texas-based AI lab Nous Research has released Arete-27B, a version of Alibaba’s open-weight Qwen3.8-27B model that has undergone a post-training process to adjust its political responses.
Nous Research stated that while Chinese open-weight models are highly capable, their underlying weights often contain state-aligned behaviors. In testing, the original Qwen model reportedly provided euphemistic responses or refused to discuss sensitive topics, such as the 1989 Tiananmen Square events.
According to internal evaluations by Nous Research, the Arete-27B model saw its factuality score increase from 2.32 to 2.98 out of 5, while responses containing political propaganda allegedly dropped from 22% to 1%. The lab noted that the propaganda layer in these models can be significantly altered through supervised learning using approximately 50 million tokens.
Entities
Alibaba · Arete-27B · Nous Research · Qwen