Nous Research's Arete-27B Cuts Qwen Propaganda From 22% to 0.9%

Nous Research has released an Apache 2.0 fine-tune of Alibaba's Qwen3.8-27B that it says cut propaganda-tagged answers on its own 803-prompt eval from 22% to 0.9%.

The model is called Arete-27B. 'We post-trained over Alibaba's Qwen3.8-27B to reinforce western values on a Chinese-trained model,' the model card reads. Training used about 52.8 million tokens at 64k context, 2 epochs, and 1,202 steps. 'We used supervised fine-tuning only. No RL, no preference optimization, and no architectural modification.' The teacher was an open-source model called Inkling, which generated 20,000 prompts across seven value categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy.

Grading was handled by a single LLM. 'An LLM judge, Grok 4.7, received the prompt, the model's answer, the required facts and a fixed rubric,' the card states, adding that the judge was not told which model wrote the answer. On a 0-to-5 factual-recall rubric scored against 803 prompts drawn from English-language government statutes, news and research reports, the base Qwen landed at 2.32, Arete at 2.98, and the Inkling teacher at 3.96.

Capability benchmarks moved little. Hermes Bench went from 65.77% to 65.60%, which Nous calls noise. Terminal-Bench-2.1 climbed from 67% to 70%.

The framing is unusually explicit for a release note, and it lands amid a steady flow of open-weights work built on top of Chinese base models that we have tracked across the past three months. 'The users of a model deserve optionality in which values they work with,' Nous writes.