Nous Research released Arete-27B in November 2026, presented as an open-weight model with a Western stance. It is not built from scratch: Nous retrained Alibaba’s Qwen3.8-27B, and in Nous’s words the aim is to reinforce Western values on a Chinese-trained model. In Nous’s test, the model repeats China’s official line less often on sensitive topics.
What is Arete?
Open weights means the model’s parameters are public, so anyone can download it, run it, and retrain it. Qwen3.8-27B is this kind of model. Nous applied fine-tuning (extra training on top of an existing model to change its behavior) to it and produced Arete-27B.
Why did Nous change it?
Nous gives several reasons in its post:
- If every AI shares the same viewpoint, that is a risk in itself. Open weights let different teams adjust a model to their own values.
- The US government restricts federal agencies from using Chinese AI models, with DeepSeek the main target.
- When Nous tested the base model itself, it saw refusals, topic changes, and answers that followed China’s official line.
How was Arete trained?
The method is simple, in three steps:
- Pick a teacher model. Nous chose Inkling, because on the topics Nous targeted it was the best of the open models with a positive test result.
- Generate about 20,000 prompts across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment.
- Have the teacher answer, then train Qwen to imitate these prompt-and-answer pairs, about 52.8 million tokens in total (tokens are the units a model reads and writes).
This method is called supervised fine-tuning: show the model examples and let it learn from them. Nous says it used only this, with no other training techniques and no change to the model architecture.
Arete’s test results
Nous built its own 803-prompt test, Propaganda_Eval. An AI judge (Grok 4.7) checks two things: whether the answer gets the facts right, and whether it adopts China’s official line.
| Model | Mean fact score (out of 5) | Share of answers with the official line |
|---|---|---|
| Qwen3.8-27B (original) | 2.32 | 22% |
| DeepSeek V4.1 Flash | 2.39 | 16% |
| GLM 5.3 | 3.10 | 15% |
| Arete-27B | 2.98 | 0.9% |
All results are with thinking mode off (the feature where a model reasons step by step before answering). Data is from the Nous post. Arete and Qwen were tested at temperature 0 (a setting that controls how random the answers are) and the other models at 0.6, so the settings are not identical.
The biggest gain was in freedom of expression: the fact score rose from 0.89 to 2.44, and the share of official-line answers fell from 60% to 1%. On general ability, Nous’s internal test shows almost no change.
How to read these results
- The prompts and the judge both come from Nous. Nous designed the test, and the judge is another AI.
- Nous draws the line for “official line.” Give the same answer to a different judge and the result may differ.
- Nous does not claim to have solved everything. The model page says it addressed “one potential root cause” of this behavior.
- Nous concludes that this layer of bias is shallow. It argues that about 50 million tokens of examples can remove this bias, and calls for models that are open, testable, and adjustable.
An open-weight model’s stance is not fixed at release. The same method can be used in other directions, so when you ask any model about a sensitive topic, look at its training background too. For how a big company approaches open weights, compare Meta Muse Spark Open.
Key takeaways
- Arete is a retrained version of Qwen, not a new model built from scratch.
- The method is to teach Qwen to imitate the teacher model’s answers.
- Nous measured the results itself: the share of official-line answers fell from 22% to 0.9%, and a different yardstick might give a different number.
If you want to try it, the weights are on Hugging Face.