Nous Research's Arete-27B cuts Qwen's state-line answers from 22% to 0.9% with SFT alone

On November 6, Nous Research released Arete-27B, a post-trained version of Alibaba’s Qwen3.8-27B that Nous describes as tuned to reinforce Western values. On Nous’s own 803-prompt evaluation, the share of answers that adopt the Chinese government’s official line fell from 22% for the base model to 0.9%, and factual accuracy on those prompts rose from 2.32 to 2.98 out of 5. General capability barely moved. The weights, the evaluation set and the training pairs are all released under Apache 2.0.

The result matters because Chinese open-weight models are among the strongest available at their sizes, and teams building on them have had to accept whatever political behavior came in the weights. Arete suggests that behavior can be largely removed with ordinary supervised fine-tuning, at a cost well within reach of a small lab.

How Nous did it

The recipe is deliberately plain. According to the model card, Arete is supervised fine-tuning only: “No RL, no preference optimization, and no architectural modification.”

• About 20,000 prompts across seven value categories (freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, environmental policy), deduplicated against the evaluation set.
• About 52.8 million training tokens at 64k context, two epochs, 1,202 steps at batch size 48, with the vision encoder frozen and the multi-token-prediction head disabled.
• Target answers came from a teacher model, Inkling from Thinking Machines Lab, which itself scored 0.3% on the propaganda measure.

The teacher choice has its own story. The Decoder reports that Nous considered Nvidia’s Nemotron 3 Ultra but dropped it after noticing irregularities on China-related topics, which Nous suspects leaked in from Chinese open models used in its training data. Nous says it has seen similar contamination in its own projects. In other words, state-aligned framing is spreading through synthetic data pipelines into models that were never trained in China, which makes measuring it more important than any single fine-tune.

The evaluation and the numbers

Each of the 803 prompts is built from a dated source, such as a statute, news report or research paper, with two to four facts an accurate answer should contain. No source supplies more than 13 prompts. xAI’s Grok 4.7 grades each answer from 0 to 5 on those facts, blind to which model wrote it, and separately flags whether the answer repeats the official narrative, endorses it, or presents it as one equally valid view. Five runs per model are reported.

Model	Fact score (0-5)	Any propaganda
Inkling (teacher)	3.96	0.3%
Mistral Large 4	3.65	0.3%
Nemotron 3 Ultra	3.22	0.6%
Arete-27B	2.98	0.9%
GLM 5.3	3.10	15%
DeepSeek V4.1 Flash	2.39	16%
Qwen3.8-27B (base)	2.32	22%

The biggest shift was in freedom of expression, where the fact score went from 0.89 to 2.44 and the propaganda rate from 60% to 1%. Answers scoring zero on facts fell from 359 to 216 of 803, and Nous says the improvement held in all seven categories.

On capabilities, Nous’s internal Hermes Bench went from 65.77% to 65.60%, which it treats as noise. Terminal-Bench 2.1 rose from 67% to 70%, but Nous does not claim credit for that, citing harness and inference differences.

The caveats are significant

Nous is upfront about most of them, and they are worth taking seriously before treating 0.9% as a settled number:

• Everything is self-reported. Nous built the evaluation, picked the judge and ran the comparisons. There is no independent rerun yet.
• One LLM judge. The rubric was refined on a 25-prompt pilot graded by hand, then applied by Grok 4.7. Whether an answer counts as “propaganda” is a judgment call, and a different judge could move these numbers.
• Mismatched decoding. Arete and the base model were run with thinking off at temperature 0; the other models in the table used temperature 0.6. At 0.6, Arete scores 2.88 and the base model 2.24 on facts, so the ranking holds, but the cross-model comparison is not apples to apples.
• Facts still trail. Arete’s 2.98 is below Mistral Large 4, Nemotron 3 Ultra and GLM 5.3. Removing the official framing is not the same as knowing the facts; a 27B model fine-tuned on 20,000 prompts still leaves many questions half-answered.
• Not a full fix. The model card says plainly: “That is not to claim that we remedied all behaviors associated with the extant viewpoint.”

There is also a framing question the release does not settle. Nous’s earlier Hermes 3 was built to be neutral and steerable to any worldview via the system prompt. Arete swaps that for a specific values target, and “Western values” is itself a position. For many teams the practical goal is factual accuracy on contested topics, which the fact score measures more directly than the propaganda flag.

Why it matters for teams building on Chinese open models

Chinese labs dominate open-weight downloads (Hugging Face reported Chinese models overtaking US ones on the Hub), and Beijing has started clearing labs for model exports under a new control regime. Nous’s claim that the alignment is “a thin layer”, applied late and removable with modest targeted supervision, has two readings. It is good news for developers who want Qwen-class capability without the framing. It also suggests the layer is just as easy to add, to any open model, by anyone with the same budget.

What to watch: independent reruns of the released Propaganda_Eval set with a different judge, whether Nous applies the same recipe to larger Qwen or DeepSeek checkpoints, and whether Alibaba’s teased Qwen4 changes how deep the alignment goes.