Editorial illustration for Qwen3.8-27B: 22% propaganda; Western models below 1%; fine-tuning strips layer
Qwen3.8-27B Has 22% Propaganda; Fine-Tuning Fixes It
Qwen3.8-27B: 22% propaganda; Western models below 1%; fine-tuning strips layer
Nous Research says it found a crack in the armor of Chinese open-weight AI models, and the fix took just 53 million tokens. The US startup, known for its open-source Hermes agent, ran Alibaba's Qwen3.8-27B through supervised fine-tuning and watched its propaganda rate fall from 22 percent to under one percent. The resulting model, Arete-27B, kept its general capabilities intact while shedding the refusals and evasive answers that showed up when the base model touched topics like free speech or democracy.
The claim matters because Chinese open-weight models have built a reputation as some of the strongest available, free to download and modify. But Nous argues that strength comes bundled with state-mandated behavior baked directly into the weights, the kind that surfaces when a model dodges a question about censorship or echoes Communist Party talking points. To test for it, Nous built a 803-prompt benchmark spanning seven categories, from surveillance and privacy to environmental policy, each one tied to facts pulled from dated sources like laws, news reports, or academic papers. The company has released both the model and the training data publicly.
Nous Research fine-tuned Alibaba's Qwen3.8-27B on 53 million tokens to strip out Chinese state propaganda. The resulting model, Arete-27B, drops its propaganda rate from 22 percent to under one percent while keeping general capabilities intact.
Why this matters
For developers building on open-weight Chinese models, Nous Research's result is actionable in a way most AI safety research isn't: 53 million tokens of supervised fine-tuning, and a 22 percent propaganda rate drops below one percent, with capabilities intact. That's a cheap fix, not a moat. It also reframes the real question for anyone evaluating Qwen, GLM, or DeepSeek for production: the bias isn't architectural, it's a training artifact sitting close to the surface.
Nemotron showing "signs of" similar issues is worth watching too, since it suggests this isn't a China-only problem but a byproduct of whatever data pipelines produce these base models. For founders weighing cost against politics when picking a foundation model, this changes the calculus; "the model is propaganda-laced" is no longer a permanent liability if a modest fine-tuning pass can strip it out. For researchers, Arete-27B is a template worth replicating across other models to see how consistent that thin-layer finding really is, and whether the same cheap intervention works on reasoning-heavy tasks, not just factual-accuracy benchmarks.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives SupportNet Uses Largest k Scores to Route Queries for Faster, Efficient RAG
From Archives Enterprises Pair Smaller Language Models With LLMs in Tiered Architectures
Common Questions Answered
What propaganda rate did Nous Research find in Alibaba's Qwen3.8-27B model?
Nous Research discovered that Alibaba's Qwen3.8-27B model had a 22 percent propaganda rate in its responses. This finding revealed a significant issue with Chinese state propaganda embedded in the base model's outputs.
How much training data did Nous Research use to reduce propaganda in Qwen3.8-27B?
Nous Research used supervised fine-tuning on just 53 million tokens to strip out the propaganda from Qwen3.8-27B. This relatively small amount of fine-tuning proved highly effective at reducing the propaganda rate from 22 percent to under one percent.
What is Arete-27B and how does it differ from the original Qwen model?
Arete-27B is the fine-tuned version of Alibaba's Qwen3.8-27B created by Nous Research after applying supervised fine-tuning. The resulting model maintains the general capabilities of the original while reducing its propaganda rate from 22 percent to under one percent and eliminating refusals and evasive answers on sensitive topics.
Why does Nous Research's finding suggest that propaganda in Chinese AI models is not an architectural problem?
Nous Research's results demonstrate that the high propaganda rate in Qwen3.8-27B is a training artifact sitting close to the surface rather than a fundamental architectural issue. The fact that just 53 million tokens of fine-tuning can reduce propaganda from 22 percent to below one percent while keeping general capabilities intact shows the bias is removable and not deeply embedded in the model's design.
What does Nous Research's work suggest about evaluating Chinese open-weight models like Qwen and DeepSeek for production use?
Nous Research's findings reframe how developers should evaluate Chinese open-weight models by showing that propaganda issues are not architectural moats but rather training artifacts that can be cheaply fixed. This means the real question for production deployment is whether the bias can be stripped through fine-tuning rather than whether the model is fundamentally unsuitable for use.
Further Reading
- State propaganda in some Chinese AI models is only “a thin layer” that fine-tuning can strip away - The Decoder
- Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values - Nous Research
- Nous Research’s Arete-27B Cuts Chinese State-Aligned Responses by 95% - AlphaSignal
- NousResearch releases Arete-27B, a Qwen3.8-27B fine-tune reducing propaganda to 0.9% - AI Briefing
