Arete-27B: Nous Research Retrains Qwen to Cut Chinese Propaganda

Nous Research has released Arete-27B, an open-weight version of Alibaba's Qwen3.8-27B that it post-trained to drop Chinese state framing on political topics. On the company's own 803-prompt test, answers containing propaganda fell from 22% to 0.9%, while general and coding-agent benchmark scores stayed roughly flat. The weights, evaluation set and training data were published on Hugging Face under an Apache 2.0 license on November 7, 2026.

The headline numbers come from Nous itself, not an independent lab, and they have not yet been replicated by outside researchers.

What is Arete-27B and how was it trained?

In its announcement post, Nous says it chose Qwen3.8-27B because it was the strongest model that runs locally and because it showed refusals, deflections and Communist Party-framed answers on sensitive subjects. The fix was plain supervised fine-tuning: about 52.8 million tokens over two epochs and 1,202 steps, with no reinforcement learning, no preference optimization and no change to the architecture.

The teacher model was Inkling from Thinking Machines Lab, picked partly because using an open model avoids terms-of-use bans on distillation. Thinking Machines has since moved on to Inkling 2 Small. Nous says it also considered Nvidia's Nemotron 3 Ultra but found that it, too, had a "checkered record" on China-related topics.

How did Nous measure Chinese propaganda in Arete-27B?

The test covers seven areas: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy and the environment. Each prompt has two to four required facts from a dated source. An LLM judge, xAI's Grok 4.7, checked whether each answer stated, missed or contradicted those facts, and whether it adopted the official Chinese line.

Model	Fact score (0-5)	Answers with propaganda
Qwen3.8-27B (base)	2.32	22%
DeepSeek V4.1 Flash	2.39	16%
GLM 5.3	3.10	15%
Arete-27B	2.98	0.9%
Mistral Large 4	3.65	0.3%

The biggest shift was in freedom of expression, where the propaganda rate fell from 60% to 1%. According to the Arete-27B model card, the model scored 65.60% on Nous's internal Hermes Bench against 65.77% for base Qwen, and 70% versus 67% on Terminal-Bench 2.1.

Why does it matter for open-source AI?

Chinese labs ship many of the strongest open models, from Qwen to Tencent's Hy4, and companies outside China increasingly run them in their own products. Nous argues that the state-aligned layer in these models is "applied late" and "applied thin," so a relatively small fine-tune can remove it without hurting capability.

There are caveats. Nous designed the categories, the rubric and the prompts, and an AI judge scored the answers. "Western values" is also a viewpoint choice rather than neutrality, which Nous acknowledges by saying the model is meant for users "where users want this change in viewpoint."

Our take

Arete-27B is less about one 27B model than about a method. If about 53 million tokens of supervised training can reshape how an open model talks about politics, the same trick works in any direction. That makes open evaluations like this one, which anyone can rerun, more important than the claims attached to them.