Arete-27B: Nous Reports 22% to 0.9% Propaganda Drop

A thin layer, removed on a home-built test

The result Nous wants readers to take away is about depth. In its words, quoted by The Decoder, the bias in a Chinese-aligned model "is applied late, it is applied thin, and fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath." Treat that sentence as what it is: Nous's hypothesis about where the behavior sits in the model, not a measurement of the model's internals. Arete is not a retrain, and Nous does not present it as one.

Relative to pretraining a 27-billion-parameter model, this was a limited post-training intervention: approximately 52.8 million tokens over two epochs, without reinforcement learning or architectural changes. Nous does not disclose the training cost, so the experiment does not establish how cheaply another organization could reproduce it.

Nous reports that Arete scored 65.60% on its internal Hermes Bench, compared with 65.77% for the Qwen base model—a decline of 0.17 percentage points that Nous interprets as noise. On Terminal-Bench 2.1, Arete scored 70% against the base model's 67%, but Nous explicitly declines to claim an improvement because configurations can differ. These results provide limited evidence against a large regression on the reported tests. They do not establish that the rest of the model's capabilities or behaviors were unchanged.

The measurement deserves the same care as the training. Nous built the evaluation, wrote the rubric, selected the dated sources each prompt's facts come from, and chose the judge. That judge is Grok 4.7, a model from xAI, grading blind — not told which model wrote each answer — with a rubric Nous revised three times against a 25-prompt hand-graded pilot. Grok 4.7 also gated the set itself: a prompt was kept only if the judge scored 4 or 5 on it, so every prompt in the evaluation is one the judge could already answer from public information. None of that makes the numbers invented. It does mean the 22% and the 0.9% are coordinates inside one lab's framework, not readings from a neutral instrument.

Nous has published the model weights, evaluation dataset and training data, making its approach open to external inspection and follow-up testing. Its model card also describes the judging rubric and evaluation design, and the evaluation release itself ships the judge prompt (judge/JUDGE_PROMPT.md) with a reference grader and scorer (judge/grade.py and judge/score.py), so anyone can rerun the test or argue with it. Publication of these materials improves auditability, but it is not independent replication, and the results still depend on Nous's rubric, its sources and its chosen judge.

Two details on the card itself should travel with the headline. First, the decoding settings are not uniform: Arete and its base model ran at temperature 0 with thinking disabled, while every other row in the comparison table, including GLM 5.3 at 15% and DeepSeek V4.1 Flash at 16%, ran at temperature 0.6. Nous prints that caveat on the card, and at temperature 0.6 Arete still leads its base model on facts, 2.88 to 2.24. Second, the fact score itself is modest: after the fix, Arete averages 2.98 out of 5. In Nous's reported results, 216 of 803 evaluated answers received a fact score of zero, down from 359 for the base model. Under Nous's rubric, contradicting even one required fact forces a zero, and unanswered prompts also score zero. The count therefore does not mean that all 216 responses were wholly fact-free or that each contained the same type of error. The card reports those counts inside its five-run comparison without stating whether they come from one run or are averaged across the five. Reducing the flagged share is not the same as answering well.

The teacher problem

Choosing what to teach required choosing a teacher, and that choice is where the project gets interesting. Nous picked Inkling, Thinking Machines' open model, as the best open model with a positive result on the evaluation it was targeting, and partly because an open teacher avoids the terms-of-use restrictions of distilling from a closed one.

Another candidate was Nvidia's Nemotron 3 Ultra. According to The Decoder's account, Nous reported irregularities on China-related topics and suspected that some behavior might reflect the use of Chinese open-weight models during training. That is Nous's interpretation of its candidate testing, not an independently established explanation of Nemotron's behavior.

The Decoder reports that Inkling's synthetic training data involved models including Kimi K2.5. That complicates any simple division between Chinese-origin and non-Chinese-origin training sources, but it does not establish that Inkling inherited the specific behaviors Nous targeted.

That circularity does not sink the result. It does set its limit. Every step, from which facts count as required to which model's answers count as the target, is a judgment call made by someone. Arete makes those calls visible and ships them as data. Earlier work pointed the same way from the outside.

The Decoder also cites earlier NewsGuard and Aleph Alpha studies reporting politically sensitive failure modes in language models. Those studies used different questions, models and evaluation criteria. They provide context for why researchers investigate these behaviors, but their percentages cannot be compared directly with Nous's 22% and 0.9% figures or treated as independent replication of Arete's results.

A viewpoint swapped is not a viewpoint removed

Read the first line of the model card again: the goal was "to reinforce western values on a Chinese-trained model." Nous is not claiming neutrality, and Arete is a deliberate break from the company's own Hermes 3 in 2024, which was built to stay neutral and adapt to whatever the system prompt asked. The argument Nous is making is about optionality: open weights let users choose which values a model ships with, instead of accepting the defaults baked in by the original lab.

That argument is worth taking seriously, and it cuts in more than one direction. Arete provides one documented example of supervised fine-tuning substantially changing a model's answers on a targeted political evaluation. Whether similar interventions would work as well on other models, other political viewpoints or independently designed tests remains unestablished.

A team running Qwen downstream now has a worked example, with data, for auditing those defaults rather than discovering them in production. It also has a reminder that the audit itself needs an author. Arete's test can tell you whether a model adopts Beijing's official account as its own view. Deciding what should replace it is not a measurement. It is the product.

Sources

[1] Nous Research, Arete-27B model card, Hugging Face (NousResearch/Arete-27B), released 6 November 2026, read in full for this article. Purpose, method, training settings, evaluation design, the full results table, the decoding caveat and the capability figures all come from this card.Read source
[2] Nous Research, Propaganda_Eval dataset and documentation, Hugging Face (NousResearch/Propaganda_Eval), read in full for this article. The rubric and scoring rules, the judge prompt (judge/JUDGE_PROMPT.md, v3), the reference grader and scorer (judge/grade.py, judge/score.py), the run-averaging rule, the answerability gate and the dataset's MIT License all come from this card.Read source
[3] The Decoder, "State propaganda in some Chinese AI models is only 'a thin layer' that fine-tuning can strip away", published 7 November 2026. The teacher-shortlist detail, the "thin layer" quotation, the Hermes 3 contrast, the NewsGuard and Aleph Alpha context, and the Apache 2.0 licence for the weights as reported there and in secondary accounts.Read source
[4] Sulat, "Nous Research's Arete-27B retrains Qwen's politics with 53 million tokens", 7 November 2026. Secondary account of the release, including the contents of the evaluation and training-data releases and the Apache 2.0 licence for the weights.Read source
[5] Nous Research, Arete-27B announcement on X, 6 November 2026, via secondary accounts (AlphaSignal, 6 November; Sulat, 7 November). Release timing and the "reinforce Western values" framing as announced. No load-bearing figure in this article rests on these accounts.

THE ALFW VERDICT

Nous reports a substantial change in how Arete-27B answers politically sensitive questions: on its own 803-prompt evaluation, answers flagged for propaganda fell from 22% to 0.9%, while the average fact score rose from 2.32 to 2.98 out of 5. The base model and Arete used matching decoding settings, but other comparison models did not. Nous designed the evaluation, selected its sources and rubric, and used Grok 4.7 as judge; the findings have not been independently replicated. Its published materials make the approach inspectable, while limited capability tests do not establish that other behaviors were preserved. Arete is explicitly trained toward Western values, not neutrality. The experiment demonstrates a substantial measured change in one model's political responses—not that bias was eliminated, factual reliability was solved or any viewpoint can be exchanged at negligible cost.