Nous Research Releases Arete-27B, Cutting China-Favoring Replies in Qwen Model

At a glance

Nous Research released Arete-27B, an AI model further trained from Alibaba's Qwen3.8-27B. The share of replies backing the Chinese government's official line dropped from 22% to 0.9%. The model is free to download and use.

Key points

1. 1

   Nous Research released Arete-27B on November 6.

2. 2

   Replies defending the Chinese government fell from 22% to 0.9%, and in freedom of expression from 60% to 1%.

3. 3

   The extra training used about 53 million tokens of answers from another AI, and the model's structure was left unchanged.

4. 4

   In the company's own capability test, the score gap between the two models was within the margin of error.

Why it matters

As AI is used more for gathering information and judgment, political bias in its answers has become an important issue. The work shows that an open model's values can be changed with additional training.

Also reported by

• 하이테크정보누스 리서치 "중국 AI 모델 선전 편향은 얇은 층"…미세조정으로 22%에서 0.9%로