Nous Fine-Tune Cuts Qwen Propaganda to 0.9%

Nous Research on Nov. 6 released Arete-27B, an open-weight fine-tune of Alibaba's Qwen3.8-27B that the U.S. lab says cut answers echoing Chinese state narratives from 22% to 0.9% on its own 803-prompt test. The lab published the weights, the evaluation and the training data in its release post, and the results have not yet been independently reproduced.

Key takeaways

• Arete-27B's fact score rose to 2.98 from 2.32 out of 5, and answers containing propaganda fell to 0.9% from 22%, Nous reports.
• The change took about 52.8 million tokens of supervised fine-tuning over two epochs, with no reinforcement learning or architecture changes.
• Nous built the test and picked the judge, Grok 4.7, so the figures are self-reported. The eval is public for anyone to rerun.

How Nous tested Qwen for Chinese state bias

The Propaganda_Eval dataset holds 803 prompts across seven categories, including freedom of expression, democracy, and data privacy and surveillance. Each prompt carries two to four required facts drawn from a source document.

An LLM judge, xAI's Grok 4.7, marked each fact as stated, missed or opposed, without being told which model wrote the answer. An answer counts as propaganda if it adopts the Chinese state's account as its own, and the judge must quote the phrase as evidence.

Decoding settings were not identical. Arete-27B and base Qwen ran at temperature 0, the others at 0.6, according to the model card. At 0.6, Arete scored 2.88 and Qwen 2.24.

How fine-tuning changed Qwen3.8-27B

Nous used Thinking Machines Lab's open Inkling as its teacher, generating 20,000 prompts and collecting Inkling's answers. The largest shift came in freedom of expression, where the fact score rose to 2.44 from 0.89 and the propaganda rate fell to 1% from 60%.

Capability held on the tests Nous ran: 65.60% on its internal Hermes Bench against 65.77% for base Qwen. Terminal-Bench-2.1 rose to 70% from 67%, a gain the lab declines to take credit for.

"It is applied late, it is applied thin, and fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath," Nous wrote of the bias layer.

Context: limits of a "Western values" model

Arete is a value swap, not a neutral model, and Nous says so. The eval measures one direction, Chinese state framing, and its scoring depends on one LLM judge tuned on a 25-prompt hand-graded pilot.

Contamination cuts both ways. Nous says it passed over Nvidia's Nemotron 3 Ultra as teacher after finding a "checkered record" on China topics, and theorizes the bias came from working with Chinese open models. The Decoder reported that Inkling's synthetic training data also drew on Kimi K2.5.

The stakes extend past chatbots. Hirundo said on Oct. 21 that Qwen3-Coder was 19% more likely to produce a hackable app for a U.S. defense contractor than for a Chinese one. The Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek products, the South China Morning Post reported in October. AI Newsway earlier covered DeepSeek V4 Pro's low-priced general availability.

What's next

The open test is the next checkpoint. Outside groups can now run the same 803 prompts on Arete or their own models, including with a different judge.

FAQ

Is Arete-27B free of Chinese bias?
Nous does not claim that. It reports 0.9% of answers still scored as propaganda on its own test, and says it has not remedied all related behaviors.

Can I download it?
Yes. The weights are on Hugging Face, along with the evaluation and training data.

References

• Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values — Nous Research (Nov. 6, 2026)
• NousResearch/Arete-27B — Hugging Face (Nov. 6, 2026)
• NousResearch/Propaganda_Eval — Hugging Face
• Chinese AI Writes Weaker Code for Beijing's Adversaries. We Fixed It. — Hirundo (Oct. 21, 2026)
• State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away — The Decoder (Nov. 7, 2026)
• US orders DeepSeek removed from federal systems, citing national security risks — South China Morning Post (Oct. 20, 2026)