Nous Research releases Arete-27B, post-training Qwen3.8-27B toward Western values
Company reports propaganda-flagged answers on an 803-prompt eval fell from 22% to 0.9% with fact scores up, Hermes Bench within noise of the base; weights and eval open.
Why it matters
Nous fine-tuned Alibaba’s Qwen3.8-27B with about 52.8 million tokens of supervised text-only data and no RL. Its Propaganda_Eval spans seven value categories including freedom of expression and democracy. The company argues Chinese open weights ship with state-aligned behavior and that a thin late bias layer can be shifted without capability loss. Weights, the 803-prompt eval and training pairs are public. The official announcement post drew heavy engagement Friday evening UTC.
Key facts
- Post-train of Qwen3.8-27B; ~52.8M SFT tokens, no RL
- Propaganda_Eval: 803 prompts, 7 value categories
- Propaganda rate 22% → 0.9% (company report)
- Fact mean 2.32 → 2.98 / 5
- Hermes Bench 65.60% vs base 65.77%
- HF weights created 2026-11-06T19:31:28Z; announce post 21:48:45Z