NEWS · MODELS · #3257
Nous Research fine-tunes Alibaba Qwen3.8-27B into Arete-27B, cuts pro‑Beijing propaganda to <1%
Nous Research supervised-fine-tuned Alibaba's Qwen3.8-27B on roughly 53 million tokens to produce Arete-27B and says the rate of responses matching Chinese state narratives fell from 22% to under 1% while general capabilities were retained; the team published the model and training data and evaluated it on a custom 803-prompt testbed. The project used Inkling as a teacher model, judged outputs with Grok 4.7, and reports the biggest improvements in free-speech prompts.
KEY POINTS
- Nous Research supervised-fine-tuned Alibaba's Qwen3.8-27B on roughly 53 million tokens to produce Arete-27B and says the rate of responses matching Chinese state narratives fell from 22% to under 1% while general capabilities were retained; the team published the model and training data and evaluated it on a custom 803-prompt testbed.
- The project used Inkling as a teacher model, judged outputs with Grok 4.7, and reports the biggest improvements in free-speech prompts.
- If political alignment in these models is a thin, late-applied layer removable by targeted supervision, it changes how firms and regulators think about mitigation, deployment, and provenance of open-weight Chinese models.
WHY IT MATTERS
If political alignment in these models is a thin, late-applied layer removable by targeted supervision, it changes how firms and regulators think about mitigation, deployment, and provenance of open-weight Chinese models.