When an open AI model you download answers a politically sensitive question, whose point of view does it carry? Nous Research set out to make that question measurable and then tried to change the answer. The result is a new open model called Arete-27B.
What is Arete-27B?
Arete-27B was created by further training Alibaba’s Qwen3.8-27B. The 27B means the model has about 27 billion parameters, a rough measure of a model’s size. Nous Research says it observed refusals, deflections and answers framed the way the Chinese Communist Party frames them on politically sensitive topics in Qwen3.8-27B, and trained the model to change those tendencies.
The training used supervised fine-tuning only. Fine-tuning means training an existing model again on new examples to change how it behaves. No reward-based reinforcement learning and no architectural changes were used. The example answers were written by another open model called Inkling, and training ran for two passes over roughly 52.8 million tokens of data. The model is published on Hugging Face under the Apache-2.0 license, and the training data is open there too.
How was the effect measured?
Nous Research built a test of 803 prompts across freedom of expression, state ideology, democracy, education, data privacy and surveillance, state versus free market, and environmental policy. For each prompt, two to four required facts were extracted from a dated source article. A separate model, Grok 4.7, graded the answers without knowing which model had written them.
Two things were measured separately. The first is facts: does the answer contain the required information? The second is framing: does the answer adopt the Chinese state’s account as its own? Quoting the official view as “the Party describes this as…” does not count as propaganda.
What do the results show?
According to the model card, the fact score rose from 2.32 to 2.98 out of five. The share of answers adopting the state’s account fell from 22 percent to 0.9 percent. The largest change was in freedom of expression, where the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60 percent to 1 percent.
There is no clear loss in general ability. On Nous’s own general-capability test, Hermes Bench, the scores are 65.77 percent and 65.60 percent, a gap the team reads as noise. On Terminal-Bench-2.1 the score went from 67 to 70, but Nous does not claim that as an improvement.
What do these results not prove?
The same team built both the test and the grading rules, and the rules were revised three times where the judge model and human grades disagreed. Arete-27B and the starting model ran at temperature 0, while the other models compared ran at 0.6, so the settings are not identical for everyone. The fact score also still trails the teacher model Inkling’s 3.96 and Mistral Large 4’s 3.65. Nous says it addressed a likely root cause, not that it fixed every tendency in the model.
The most important thing I take from this is that relatively small training, about 52.8 million tokens, could shift a model’s stance on political topics this much. That cuts both ways: if reducing one tendency is easy, adding another may be just as easy. So when choosing a model, running the openly published 803-prompt test on the topics you care about is a more reliable guide than the headline “propaganda rate down.” Less state framing also does not mean more knowledge; a fact score that is still in the middle of the range shows that.
