Arete-27B
Nous Research released Arete-27B, Alibaba's Qwen3.8-27B post-trained to reinforce Western values, cutting the share of answers repeating China's party line from 22% to 1% on its own 803-prompt eval.
Nous fine-tuned Qwen3.8-27B on roughly 53 million tokens so it adopts China's official line less often. It built an 803-prompt eval across seven value categories, including freedom of expression, democracy and privacy. Nous says the bias layer in the Chinese model is shallow and its Hermes Bench score shows no significant performance tradeoff.
- Date
- Friday 6 November 2026
- Lab
- Nous Research
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Propaganda rate (share of answers with at least one point of party line) | 22% to 1% (0.9% in the post's table, reasoning off) 803 prompts, seven value categories | company |
| Fine-tuning data | about 53 million tokens | company |
Weights, eval and training data are reported public on Hugging Face under Apache 2.0, per secondary coverage. The results come from Nous's own eval.
Sources
- pasqualepillitteri.it/en/news/31090/arete-27b-nous-research-qwen-western-values
- techmeme.com/261108/p3#a261108p3
- xenospectrum.com/en/nous-arete-model-evaluation/
- tokenpost.com/news/technology/51818
This record was partly confirmed: some claims could not be checked on 8 November 2026. How we check