Nous Research releases Arete-27B, a Qwen fine-tune trained to drop state-aligned answers
Nous post-trained Alibaba's Qwen3.8-27B with supervised fine-tuning, stating a goal of reinforcing Western values. On its own 803-prompt test, answers containing Chinese state propaganda fell from 22 percent to 0.9 percent, and the fact score rose from 2.32 to 2.98 out of 5. Nous reports no meaningful loss on its general and agentic tests. The Decoder reports that Grok 4.7 served as the judge.
Why it matters: It shows how far post-training alone can shift a model's political behavior, though the test and its choice of values are the developer's own.
- Parameters
- 27B
- Active per token
- Not disclosed
- Weights size
- 54.7 GB (Hugging Face)
- License
- Apache 2.0
- Starting model
- Qwen3.8-27B
| Benchmark | Score | Type | Measured or published by |
|---|---|---|---|
| Propaganda rate, Nous eval | 0.9% (was 22%) | Self-reported | Nous Research |
| Fact score, Nous eval | 2.98 of 5 (was 2.32) | Self-reported | Nous Research |