What changed
Nous Research posted model files for Arete-27B on November 6. It started with Alibaba's Qwen3.8-27B and used fine-tuning, which means training an existing model on more examples to change how it responds. Nous says it wanted the model to reflect what it calls Western values and to answer politically sensitive questions differently. [1] [2]
What the evaluation tells us
Nous tested the new model on 803 prompts about subjects such as free expression, privacy, and the environment. Its evaluation scored factual recall separately from framing, or how an answer presents a subject and whose viewpoint it adopts. Nous reported higher factual scores and less use of the Chinese government's framing in its test. Another AI model graded the responses using rules set by Nous. Those results describe how the models answered that set of questions under those rules. [1] [2] [3]
Our take
When two models give different accounts of the same issue, look for a claim you can check. If you ask about a city's public camera policy, ask how long the city keeps footage and what exceptions it lists. Then open the city's policy and check those details. If an answer calls the program effective, ask what evidence supports that judgment. This helps you separate a verifiable detail from the way the model tells the story.