Nous Research fine-tuned China's Qwen3.8, and propaganda in its answers dropped from 22% to 1%
Nous Research (USA) fine-tuned Alibaba's open Chinese model Qwen3.8-27B and called the result Arete: the share of propaganda-laden answers dropped from 22% to 1%. It took 50 million tokens of data, which is modest by AI standards. Factual accuracy rose from 2.32 to 2.98 points out of five. The ideological layer in the Chinese weights turned out to be shallow, which means checking and correcting such a layer can be done cheaply.
AI-processed from 3DNews AI; edited by Hamidun News
Imagine you have built an open Chinese model into a chatbot, and on questions about politics and history it starts repeating Beijing's official position. That is the problem the American lab Nous Research set out to solve. It took Alibaba's Qwen3.8-27B, fine-tuned it on a small dataset, and produced a version called Arete. The share of propaganda-laden answers fell from 22% to 1%, and the factual accuracy score rose from 2.32 to 2.98 on a five-point scale. The whole effort used 50 million tokens of training data.
What exactly was done to Qwen3.8-27B
The developers stated the goal plainly: to anchor Western values in a model originally trained in China. Their argument is that Chinese open-weight models are now among the most capable available, yet their weights encode behavior that reflects the Chinese state ideology.
Technically, this is additional training. A useful comparison is an employee who knows their job well but answers contentious questions with memorized lines. They get a short refresher course, their knowledge stays in place, and the way they handle sensitive questions changes.
The numbers are as follows. An "answer with propaganda" is one that repeats the positions of Chinese ideology. Arete-27B gives such answers 1% of the time, down from 22%. According to 3DNews, Western models that Nous Research tested in its lab show roughly similar figures. Factual accuracy, meaning how closely answers match verifiable facts, rose from 2.32 to 2.98 out of five. The gain is noticeable, but a score just under three is hard to call flawless.
Why 50 million tokens was enough
The key observation is that the propaganda layer in Qwen3.8-27B was shallow. It was removed with 50 million tokens of data. For modern models that is a small amount, and this was light adjustment rather than training from scratch.
The choice of tools is also telling. The researchers first considered Nvidia's Nemotron 3 Ultra, but found that it too answers China-related topics ambiguously. They suspected Chinese models may have been used in training Nemotron itself. In the end they chose the open Inkling model from the American company Thinking Machines Lab. The origin of training data shapes a model's behavior, and a company name on a license does not reveal it.
Hirundo has a similar track record. It applied retraining methods to Qwen3.6-35B-A3B, and the share of answers reflecting Beijing's preferences dropped from 89.8% to 2.8%. Hirundo's effect is stronger than Nous Research's, but according to the description the two projects measure different things, so the results cannot be compared directly.
What developers should do
Risks from Chinese models are not limited to ideology. According to 3DNews, under certain conditions some of them can generate insecure program code. For that reason, such weights should not be evaluated on leaderboard rankings alone.
Practical takeaways:
- If you adopt an open Chinese model for a bot, support, or analytics, run it against a set of questions on politics, history, and contested events before release.
- For critical use cases, consider light fine-tuning: judging by these cases, it is inexpensive and works quickly.
- For code and agents, a separate security audit is needed. Clean answers on politics say nothing about how safe the generated code is.
There are caveats too. The figures come from 3DNews coverage, and the full methodology for measuring propaganda and factual accuracy is not disclosed in the text. The notion of "Western values" is also contested: it is the project authors' choice, not a neutral standard. Still, the core fact matters to anyone selecting open weights for production: an ideological overlay can be removed with a small amount of data, which means it can be measured and corrected when needed.
- Nov 21, 2026Chinese Tech Giants Double AI Spending to $139B, While Labs Recover Less Than 8%
- Nov 21, 2026SGLang with Humming Kernels Cut Time to First Token by 7–23% and Output Throughput by 17–34%
- Nov 21, 2026CoreWeave Owes $47 Billion on Nvidia Chips and Lost $709 Million in One Quarter
- Nov 21, 2026A local phi3 was jailbroken in 51% of attempts — the developer put a security gateway in front of it
Key Questions Answered
What exactly was done to Qwen3.8-27B?
The developers stated the goal plainly: to anchor Western values in a model originally trained in China. Their argument is that Chinese open-weight models are now among the most capable available, yet their weights encode behavior that reflects the Chinese state ideology. Technically, this is additional training. A useful comparison is an employee who knows their job well but answers contentious questions with memorized lines.
Why 50 million tokens was enough?
The key observation is that the propaganda layer in Qwen3.8-27B was shallow. It was removed with 50 million tokens of data. For modern models that is a small amount, and this was light adjustment rather than training from scratch. The choice of tools is also telling. The researchers first considered Nvidia's Nemotron 3 Ultra, but found that it too answers China-related topics ambiguously. They suspected Chinese models may have been used in training Nemotron itself.
What developers should do?
Risks from Chinese models are not limited to ideology. According to 3DNews, under certain conditions some of them can generate insecure program code. For that reason, such weights should not be evaluated on leaderboard rankings alone. Practical takeaways: If you adopt an open Chinese model for a bot, support, or analytics, run it against a set of questions on politics, history, and contested events before release.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.