- Nous Research fine-tuned Alibaba’s Qwen3.8-27B on about 52.8 million tokens and called the result Arete-27B.
- On an 803-question test, answers with propaganda fell from 22% to about 1%, and Nous says the model’s general skills stayed the same.
In short: Nous Research took one of the strongest free AI models, Alibaba’s Chinese Qwen3.8-27B, and used a short extra round of training to stop it repeating Beijing’s official line. The new model is called Arete-27B, and its weights, training data and test set are public on Hugging Face.
What changed
- Nous built its own test: 803 prompts across seven categories, including freedom of expression, surveillance, democracy and the environment. Each prompt comes with 2–4 required facts from a dated source article.
- Another AI model, Grok 4.7, graded the answers as a judge model. An answer counted as propaganda when it replaced facts with the official narrative or endorsed it, and the judge had to quote the exact phrase.
- Base Qwen gave propaganda in 22% of answers; Arete-27B in about 1%. The average fact score rose from 2.32 to 2.98 out of 5.
- The biggest shift was in freedom of expression, where the propaganda rate fell from 60% to 1%.
- The whole thing was fine-tuning on about 52.8 million tokens over two passes through the data, with no change to the model’s architecture.
What it costs and what happened to skills
Nous says the cost of this training is insignificant next to the cost of building the model in the first place. On its internal general-skills test, Hermes Bench, Arete-27B scores 65.60% against 65.77% for base Qwen, a gap Nous calls indistinguishable from noise. The example answers it learned from were written by another open model, Inkling, chosen as the teacher.
How to read it
These are the author’s own results, on a test the author built, with a single AI model doing the grading. On the plus side, the test, data and model are all public, so anyone can rerun it. The part that matters most to me is Nous’s conclusion: the layer of views in a model is thin, and one round of fine-tuning moves it. That means the values of any open model, Western ones included, can be pushed in whatever direction someone wants. I cover models like this in the Open models hub.