Nous Gutted Qwen's Propaganda Script With 53 Million Tokens
Nous Research fine-tuned Alibaba's Qwen3.8-27B into Arete-27B, cutting CCP-aligned answers from 22 percent to 1 percent on its own eval, with benchmark scores barely moving. The political layer was never load-bearing.

You download a Chinese open-weight model because it is free, it is strong, and nobody at Alibaba can see your prompts. Then you ask it about Tiananmen Square, or Taiwan, or who runs Hong Kong, and it answers like a state broadcaster. The assumption has always been that this is the price of admission: take the capability, eat the propaganda, or go pay OpenAI instead.
Nous Research just showed that assumption is wrong.
The lab took Alibaba's Qwen 3.8-27B and fine-tuned it into something called Arete-27B. For each prompt they extracted facts from a dated source article, along with the official-frame language the CCP uses on that topic, and an LLM judge graded answers against a fixed rubric. Arete-27B raised its fact score from 2.32 to 2.98 out of 5 and cut answers containing propaganda from 22% to 1%, in line with Western models tested.
The cost of that shift was not a retrain. Nous fine-tuned Qwen on about 52.8 million tokens, roughly fifty thousand samples at 64k context, over two epochs and 1,202 steps, with no RL, no preference optimization, and no architectural modification. For a model whose base pretraining ran on trillions of tokens, 53 million is a rounding error.
The part that should worry Alibaba
Here is the number that matters more than the censorship drop: capability didn't move. On Nous's internal Hermes Bench, Arete-27B hits 65.60 percent compared to 65.77 percent for the base model. That is not a trade. You are not giving up reasoning ability to get an uncensored model. You are just removing a layer that was never connected to the reasoning in the first place.
The freedom-of-expression category breakdown makes the point sharper. The fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%. A model that gave the party line on free speech six times out of ten now gives it once in a hundred, and it still codes, still reasons, still uses tools the same as before.
Nous's own framing is the quote to sit with: the bias layer in a Chinese-aligned model is shallow, applied late, applied thin, and fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath, because the values in any open model are only ever one fine-tune away.
Translation: whatever Alibaba thinks it baked into the weights, it's wallpaper. Anyone with a few GPUs and an afternoon can strip it and repaper the room in whatever values they want, Western, Beijing's, or someone else's entirely.

Not peer reviewed, but reproducible
Before you treat this as settled science: the headline claim is credible as a measured result on Nous's own test, but it is not yet independently confirmed. The judge model was xAI's Grok 4.7, the rubric was Nous's own design, and the comparison set was Nous's own choice of Western models. The release includes the weights, evaluation prompts, and roughly 53 million tokens of training data, which means any skeptical grad student can run the same eval and check the math.
This isn't Nous's first rodeo with Chinese open models. The lab closed a Series B in October at $1.5 billion valuation, with Robot Ventures leading and Nvidia and Samsung among participants. Stripping the geopolitics out of the strongest open model on the market is a very good pitch to enterprise customers who were told to stay away from Chinese weights entirely.
If you are running Qwen, DeepSeek, or GLM locally and treating the censorship as a fixed cost of doing business, stop. It is a 53-million-token problem, and somebody just published the recipe.
What to watch: whether an independent lab reruns the Arete eval with a different judge model and gets the same 1 percent; whether Alibaba or other Chinese labs respond with licensing restrictions instead of weight-level fixes; and whether the same recipe gets applied to DeepSeek or GLM, where nobody has published a comparable before-and-after yet.
- Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values
- State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away
- Arete-27B: Nous Retrains Qwen3.8, Claims 22% to 0.9%
- Nous Research's Arete-27B Cuts Chinese State-Aligned Responses by 95%
- Nous Research's Arete-27B Cuts Chinese State-Aligned Responses by 95%
- Nous Research launches Arete-27B and cuts Qwen's propaganda rate from 22% to 1%
- Nous Research on X: "Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of ex… / X
- Nous Research's Arete-27B Cuts Qwen Propaganda From 22% to 0.9%
- Nous Research unveils Arete-27B, a post-trained model built to reduce Chinese AI propaganda
- Arete-27B: Nous Retrains Qwen to Cut Chinese Propaganda
- Nous Research's Arete-27B Strips Chinese State Propaganda From Qwen Down to 0.9%
- Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9% - MarkTechPost
- Luke The Dev on X: "Arete-27B is live Nous took Qwen3.8-27B and cut CCP-framed answers from 22% to 1% on an 803-prompt values eval. Hermes Bench barely moved: - Qwen3.8-27B: 65.77% - Arete-27B: 65.60% 53M tokens of SFT. No RL. Same agentic score. If you run local agents on Chinese open weights,… / X
- FHILY👑 on X: "Arete-27B takes Qwen3.8-27B and fine-tunes its responses around a different set of political and social values. Developed by Nous Research, the model explores an important feature of open-weight AI: a model’s behaviour is not necessarily fixed by the preferences present in its origi… / X