Nous Research releases Arete-27B, a Qwen3.8-27B fine-tune that cuts CCP-aligned answers from 22% to ~1% on its own eval
Confirmed
The takeaway
Nous Research post-trained Alibaba’s Qwen3.8-27B on ~53M tokens to remove Chinese state-aligned behaviour; on its 803-prompt eval (graded by an LLM judge) the ‘propaganda rate’ fell from 22% to about 1%, with no significant loss on its Hermes Bench.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 8 November 2026
Your AI and this story
- GPT-6 Astra190 days after its cutoff
- Claude Opus 5.5129 days after its cutoff
- Gemini 4 Argon220 days after its cutoff
- Grok 4.7159 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 129 days before it.
Key facts
- Eval: 803 prompts in seven value categories (freedom of expression, democracy, privacy…), scored by Grok 4.7 as judge
- Propaganda rate 22% → ~1% (0.9% thinking off; thinking-on score provisional, 2 of 5 runs); fact score 2.32 → 2.98 of 5
- Training: ~52.8M tokens, ~50k samples, two epochs of SFT; no RL or preference optimisation
- Hermes Bench within 0.2 points of base; Terminal-Bench 2.1 70% vs 67% (different configs, per Nous)
- Self-evaluated; Nous says not all base-model viewpoints were removed. The viral claim of ‘zero’ propaganda is wrong: ~1%
- X announcement: ~1.9M views by Nov 8
What happened
Announced Nov 6, 2026 on X and nousresearch.com/arete. Nous: “We post-trained over Alibaba’s Qwen3.8-27B to reinforce Western values on a Chinese-trained model.” Techmeme carried it on Nov 8; commentators (e.g. Curtis Yarvin) debated the framing of ‘Western values’.
Why it matters
Shows how cheaply the political behaviour of strong Chinese open-weights models can be changed, in either direction.
Sources
4 sources from 3 sites. Numbers match the chips in the text.
4 sources: 4 primary
Primary
- Nous Research: Aretenousresearch.com, official
- Nous Research on Xx.com, official
- Weights on Hugging Facehuggingface.co, code
- Propaganda eval datasethuggingface.co, code
Changes
- Filed