Editorial illustration for Nous Releases Arete-27B, Cuts Qwen3.8-27B Propaganda From 22% to 0.9%
Nous Arete-27B Cuts AI Propaganda From 22% to 0.9%
Nous Research put out Arete-27B this week, a fine-tuned version of Alibaba's Qwen3.8-27B built to strip out the state-aligned framing that shows up when the base model answers questions about China. The numbers are stark: on Nous's own 803-prompt evaluation set, the rate of answers containing propaganda dropped from 22% to 0.9%. Nothing else changed architecturally. Nous used supervised fine-tuning only, no reinforcement learning, no preference optimization, training on roughly 52.8 million tokens at 64k context with Inkling, an open model from Thinking Machines, serving as the teacher.
The choice of teacher model matters here. Nous picked an open one specifically to sidestep terms-of-use restrictions that come with distilling from closed systems, and it passed on Nvidia's Nemotron 3 Ultra after finding what it describes as a mixed record on China-related topics. The resulting model, weights, eval set and training data, is all public under Apache 2.0, the same license as the base Qwen model. Nous says it picked Qwen3.8-27B to begin with because it was the strongest local model showing the exact behavior they wanted to fix: refusals, deflections and CCP-aligned framing on sensitive subjects.
On Nous’s own 803-prompt eval, answers containing propaganda fell from 22% to 0.9%. Developers get Qwen-class agentic skill without state-aligned answers. The weights, eval and training data are all public.
Why this matters
Arete-27B is a reminder that open weights don't mean neutral weights. Qwen3.8-27B is genuinely capable, which is exactly why a 22% propaganda rate in Nous's own eval is worth noting rather than dismissing as a China-specific quirk. Any base model trained under a particular regulatory or cultural regime will carry some of that framing into its outputs, and most developers never run the eval that would show them this.
What's useful here isn't just the fix, it's the method: Nous published the training data, the eval set, and picked an open teacher model (Inkling) instead of a black-box one. That's the part worth copying. If you're building on a foundation model from any jurisdiction, you can now check for baked-in state framing yourself rather than taking a vendor's word for it.
The 0.9% residual rate is still not zero, and a single 803-prompt benchmark from the same team that built the fix is a narrow way to measure "bias." Treat this as a useful patch and a template for auditing, not a finished answer to the problem.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives AI auditors detect 80% of agent misbehavior but trigger false alarms
From Archives New report on Anthropic automating AI development alarms researchers
Common Questions Answered
What specific improvement did Nous achieve with Arete-27B compared to the base Qwen3.8-27B model?
Nous Research successfully reduced the propaganda rate in answers about China from 22% to 0.9% on their 803-prompt evaluation set. This dramatic improvement was achieved through supervised fine-tuning alone, without any architectural changes, reinforcement learning, or preference optimization techniques.
What training method did Nous use to create Arete-27B and how much data was involved?
Nous used supervised fine-tuning as the sole training methodology to remove state-aligned framing from the base model. The fine-tuning process involved approximately 52.8 million tokens of training data.
Why does the article emphasize that 'open weights don't mean neutral weights'?
The article highlights that any base model trained under a particular regulatory or cultural regime will carry framing from that context into its outputs, as demonstrated by Qwen3.8-27B's 22% propaganda rate. Most developers never run evaluations that would reveal these biases, making Nous's transparent approach and public release of weights, eval data, and training data particularly valuable.
What resources did Nous make publicly available with the Arete-27B release?
Nous released the model weights, evaluation data, and training data publicly, allowing other developers to understand the fine-tuning methodology and verify the propaganda reduction results. This transparency enables developers to access Qwen-class agentic capabilities without state-aligned answers.
Further Reading
- State propaganda in some Chinese AI models is only “a thin layer” that fine-tuning can strip away - The Decoder
- Nous Research's Arete-27B Cuts Qwen Propaganda From 22% to 0.9% - AI Weekly
- Nous Research cuts pro-Beijing answers in Alibaba's Qwen from 22% to under 1% with fine-tuning alone - MLQ.ai
- Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values - Nous Research
