Nous Research Retrains Alibaba's Qwen Model as Arete-27B

Official-narrative answers fall from 22% to 0.9%

To measure the shift, Nous built an 803-question evaluation across seven categories: free speech, state ideology, free markets, education, data privacy and surveillance, democracy, and environment. In the company's tests, the share of Qwen3.8-27B answers adopting the Chinese government's official line as its own was 22%. For Arete-27B it was 0.9%. The 5-point "fact score" rose from 2.32 to 2.98.

The biggest change came in free speech, where the fact score went from 0.89 to 2.44 and official-narrative answers dropped from 60% to 1%.

There are caveats. The evaluation was designed by Nous itself, used Grok 4.7 as the scorer, and was checked by humans on only 25 questions. No independent third party has verified it. A lower rate of official-line answers also does not mean the model is politically neutral or that everything it says is accurate.

General performance holds

On the in-house Hermes Bench, Qwen3.8-27B scored 65.77% and Arete-27B 65.60%, a gap Nous attributes to measurement noise. On Terminal-Bench-2.1, which tests agentic task completion, Arete-27B reached 70% against 67% for the base model, though the company does not credit its training method because test configurations differed.

The teacher model was Inkling. Nous generated roughly 20,000 prompts across the seven categories and built a final set of 28,886 training examples from the teacher's answers. The weights, training data and evaluation set are published on Hugging Face so researchers can test the results independently.

Nous argues the main finding is that values embedded in open-weight models can be shifted substantially with relatively little extra training.