Nous Research · 2026-11-06 · notable
Arete-27B — Nous Research strips state propaganda from Qwen3.8-27B
In short
Arete-27B is Nous Research's open-weight fine-tune of Qwen3.8-27B that cuts state-aligned answers on political topics from 22% to 0.9% on its 803-prompt test, with no loss on Terminal-Bench-2.1.
Nous Research · 2026-11-06 · 2 sources checked
An open 27B model tuned to answer politically sensitive questions with facts instead of state talking points.
Key specs
| Propaganda answers | 22% → 0.9% |
|---|---|
| Terminal bench 2.1 | 70% |
What is it?
Arete-27B is Alibaba's Qwen3.8-27B after supervised fine-tuning by Nous Research to change how it answers politically sensitive questions. The team targets seven value areas, including freedom of expression, state ideology, democracy and surveillance, where it saw refusals, deflections and Chinese-government framing in the base model. Weights, the training pairs and the 803-prompt Propaganda_Eval test set are all on Hugging Face, and the model is licensed Apache-2.0.
How does it work?
Training was supervised fine-tuning only: no RL, no preference optimization and no architecture change. Inkling was the teacher model, and about 52.8M tokens of examples were trained for two epochs at 64k context, with the vision encoder frozen. An LLM judge scores each answer for factual recall (0–5) and for propaganda framing.
Why does it matter?
Many of the strongest open models come from Chinese labs, and Nous argues their political bias is shallow enough to remove with a small fine-tune. On its own test the fact score rose from 2.32 to 2.98 and freedom-of-expression propaganda fell from 60% to 1%, while Hermes Bench stayed flat (65.77% vs 65.60%) and Terminal-Bench-2.1 rose from 67% to 70%. The authors note these runs used their own setup and that not every behavior is removed.
Who is it for?
teams that want a Qwen-class open model without state-aligned answers
Try it
huggingface.co/NousResearch/Arete-27B