Wire flash
NousResearch releases Arete-27B, fine-tuned Qwen model to reinforce Western values
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
NousResearch announced Arete-27B, a post-trained version of Qwen3.8-27B designed to reinforce Western values. The organization states that Chinese open models are currently the strongest available but ship with state-aligned behavior in their weights. To measure this, NousResearch built an 803-prompt evaluation across seven value categories, including freedom of expression, democracy, and privacy, tracking how often a model adopts China's party line. After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models tested, according to the post. Its Hermes Bench score shows no significant performance tradeoffs. The announcement concludes that the values in an open model are one fine-tune away. The evaluation, weights, and training data are available now alongside a full blog post.
Source report
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China's party line.
Key Results
- After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested.
- Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Availability
Our eval, weights, and training data are available now, along with the full blog post: https://t.co/iAOANQTVZ6
Source
NousResearchNeutral / independent
Part of this Story
NousResearch Releases Arete-27B, Cuts Flagged Propaganda From 22% to 1%