Nous Research has unveiled Arete-27B, a version of Alibaba's Qwen3.8-27B that the US lab post-trained to reinforce Western values. The company announced the model in a blog post dated November 2026, describing the work as an attempt to reshape the behavior of a Chinese-trained open-weight model.
The move lands in the middle of a heated debate over how much political and ideological framing Chinese AI models carry, and whether foreign weights should be allowed into Western companies and government agencies. Nous says it measured exactly what changed, using an evaluation of 803 prompts split across seven value categories.
What Arete-27B measures
Nous built the benchmark before training anything, so it could define what it was aiming at. Each of the 803 prompts carries two to four required facts drawn from a dated source article, plus the official framing language used on that topic. Prompts came from English-language government statutes, news and research reports, and no single source supplied more than 13 prompts.
An LLM judge, Grok 4.7, received the prompt, the model's answer, the required facts and a fixed rubric, and was not told which model produced the answer. For every fact it recorded whether the answer stated, missed or contradicted it, and it had to quote an actual phrase as evidence before a propaganda score could be counted. Nous says it developed the judge on a 25-prompt pilot it graded by hand, revising the rubric three times when machine and human grades disagreed.
- Base model: Alibaba's Qwen3.8-27B, an open-weight Chinese model
- Evaluation set: 803 prompts across seven value categories
- Categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment
- Two axes measured: factual recall and framing (including identification of official CCP-aligned language)
- Judge: Grok 4.7, scoring answers without being told which model wrote them
