Nous Research has released Arete-27B, an open-weights language model built by retraining Alibaba's Qwen3.8-27B to reduce answers that repeat the Chinese state's official line on politically sensitive questions. The lab announced the model on 6 November and published the weights, training data and evaluation set together.

On Nous's own 803-prompt evaluation, called Propaganda_Eval, the share of answers adopting the official line fell from 22 percent for the base Qwen model to 0.9 percent for Arete-27B. The lab also reports that factual accuracy on those prompts rose from 2.32 to 2.98 out of 5, while general capability barely moved.

How the model was retrained

The method is deliberately simple. Nous used supervised fine-tuning only, on about 52.8 million tokens, with no reinforcement learning, no preference optimisation and no change to the model architecture. The training data covered about 20,000 prompts across seven areas, including freedom of expression, state ideology, democracy, education and data privacy. Answers came from a teacher model called Inkling, from Thinking Machines Lab.

Open release, with limits

The weights are available on Hugging Face under the Apache 2.0 licence, the same licence Alibaba uses for Qwen3.8-27B. Nous says the release is the first to publish the full realignment recipe as open weights, so other researchers can download the model and rerun the same test.

The figures should be treated with care. The evaluation was built and run by Nous itself, judged by Grok 4.7, and has not yet been independently rerun. Nous states plainly that the results do not claim to have remedied all behaviours tied to the original viewpoint.