Nous Research has released Arete-27B, a retrained version of Alibaba's Qwen3.8-27B that answers politically sensitive questions differently from the model it started from. On the lab's 803-question test, the share of answers repeating Beijing's official line fell from 22 percent to 0.9 percent, while general ability barely moved.
The result lands amid a live debate over open-weight AI. Chinese open models are among the strongest available at their size, and developers building on them have had to accept whatever political behaviour came baked into the weights. Arete suggests that behaviour can be stripped out with an ordinary fine-tuning run.
A deliberately simple recipe
Nous says the method is deliberately plain. Arete is supervised fine-tuning only: no reinforcement learning, no preference optimisation, and no change to the model's architecture. The lab ran about 53 million tokens of training over two epochs and 1,202 steps, from roughly 20,000 prompts across seven value categories — freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy.
The target answers came from Inkling, a teacher model from Thinking Machines Lab that itself scored 0.3 percent on the propaganda measure. The weights, training data, and evaluation set are all public on Hugging Face under the Apache 2.0 licence — the same licence Alibaba uses for Qwen.
How the test works — and its limits
Each of the 803 prompts is built from a dated source — a statute, news report, or research paper — and carries two to four facts an accurate answer should contain. xAI's Grok 4.7 grades each answer from zero to five on those facts, without knowing which model wrote it, and separately flags whether the answer repeats the official narrative, endorses it, or presents it as one equally valid view.
The headline numbers are striking. The fact score rose from 2.32 to 2.98 out of five, with the biggest shift in freedom of expression, the score climbing from 0.89 to 2.44, the propaganda rate dropping from 60 percent to 1 percent. For comparison, the unmodified Qwen scored 22 percent, DeepSeek's V4.1 Flash 16 percent, and GLM-5.3 15 percent, while Western open models all landed under 1 percent.
Nous is upfront about the caveats. The evaluation is self-reported — the lab built the test, picked the judge, and ran the comparisons, with no independent rerun yet. A single model, Grok 4.7, did all the grading, and whether an answer counts as propaganda is a judgment call. And the facts still trail: Arete's 2.98 sits below Mistral Large 4, Nvidia's Nemotron 3 Ultra, and GLM-5.3. The model card says plainly that the team is not claiming to have fixed every behaviour tied to the original model's outlook.
Why it matters
Even so, the release reframes a question the industry has treated as settled. If a political layer in a strong open model is shallow — applied late, applied thin, and removable with about fifty million tokens of careful supervision — then the values baked into any open model are, in Nous's words, "one fine-tune away." That cuts both ways: the same technique could strip out one set of views or install another, which is why the lab published the training pairs alongside the weights.
For now, Arete-27B is a proof of concept rather than a product. No inference service hosts it yet, so running it needs a large GPU or a community quantization. But as a demonstration that open models can be re-aligned after the fact — cheaply and openly, without losing capability — it is one of the more consequential releases of the month.
