Nous Research releases Arete 27B to cut state framing down to under one percent
Executive Briefing
Synthesized and verified with gemini-aistudio-api-key/models/gemini-3.5-flash-lite / reviewer requested: gemini-aistudio-api-key/models/gemini-3.5-flash
“Nous Research has released Arete 27B to significantly reduce state-aligned messaging...”
Key Takeaways
- 1.Arete 27B is a supervised fine-tune based on Alibaba's Qwen3.8-27B model.
- 2.On the research team's 803-prompt evaluation, answers containing propaganda fell from 22% to 0.9%.
- 3.The training used only supervised fine-tuning over about 52.8 million tokens without reinforcement learning or preference optimization.
Nous Research has released Arete 27B to significantly reduce state-aligned messaging inherited from its base model. The fine-tuning process preserved core model capabilities without relying on reinforcement learning.
Key Developments and Technical Context
Arete 27B is a supervised fine-tune based on Alibaba's Qwen3.8-27B model.
On the research team's 803-prompt evaluation, answers containing propaganda fell from 22% to 0.9%.
The training used only supervised fine-tuning over about 52.8 million tokens without reinforcement learning or preference optimization.
Strategic Analysis and Industry Implications
Direct supervised fine-tuning using open teacher data demonstrates that the community can adjust the geopolitical framing of large open models without degrading overall technical capabilities.
Outlook and Key Factors to Watch
It remains unclear how the model performs on intensive domain benchmarks beyond the evaluation sets published by the authors.
