Nous says a 53M-token fine-tune stripped Qwen3.8-27B's political slant; Arete-27B released as open weights
According to The Decoder on November 7, 2026, Nous Research supervised-fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens to produce the open-weights model Arete-27B. On Nous's own 803-prompt test, propagandistic answers fell from 22% to 0.9% and factual accuracy rose from 2.32 to 2.98, with general capability said to be largely unchanged.

What changed
The core of the report is not a new base model but a post-training experiment targeting alignment behavior. According to The Decoder's November 7, 2026 article and Nous Research's account, Alibaba's Qwen3.8-27B refuses, deflects and echoes official positions on sensitive topics; Nous trained Arete-27B with about 53 million tokens of supervised fine-tuning over two epochs, using no reinforcement learning or preference optimization, and says it removed overlap between training and test prompts beforehand. The evaluation is Nous's own: 803 prompts in seven categories covering free speech, state ideology, education, privacy and surveillance, democracy and the environment. Each prompt comes with 2 to 4 facts from dated sources (laws, news reports or research papers), and xAI's Grok 4.7 serves as judge, deciding whether an answer mentions, omits or contradicts the facts; the judge must quote specific wording and does not know which model produced the answer. The report says unmodified Qwen3.8-27B scored 2.32 out of 5 on factual accuracy with 22% of answers containing propaganda; Arete-27B scored 2.98 and 0.9% respectively. The biggest drop was in the free-speech category, where the factual score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%. In the same comparison, the report puts GLM-5.3 at 15% and DeepSeek V4.1 Flash at 16%, with Western models such as Mistral Large 4 and Gemma 4 31B below 1%; it also cites ValsAI research finding that different Chinese models behave inconsistently. On general capability, Nous's internal Hermes Bench score is 65.60% versus 65.77% for the base model; Terminal-Bench 2.1 rose from 67% to 70%, though Nous itself says the test setup differed and this is not a real improvement. Nous concludes that the bias layer in models aligned with Chinese positions is "shallow": applied late and thinly, removable with fifty million tokens of curated supervision without disturbing the layers beneath. The report says Arete-27B's weights, test dataset and training data are public on Hugging Face under the Apache 2.0 license, and that Nous suspects anomalies in Nvidia's Nemotron 3 Ultra on China-related topics stem from contamination by training data generated with Chinese open-weights models, a problem it says its own project also encountered.
What it means for developers
The following is analysis and judgment based on the above report, not a reproduction of its results by this site. The first thing this news changes is expectations about how portable the "alignment layer" is. If Nous's method and numbers hold up, resetting how a Chinese open-weights model answers on sensitive topics may cost on the order of tens of millions of tokens of supervised data plus one routine SFT run, rather than retraining from scratch or relying on reinforcement learning. That directly affects whether you would put such weights into settings with requirements about how positions are expressed, such as multi-region content moderation assistance, policy retrieval-augmented generation, or public-opinion summarization. Conversely, the report also notes that bias may pass between student models through synthetic data: Nous says it saw anomalies in Nemotron 3 Ultra, suspects the source is Chinese open-weights models in the training data, and says its own project hit similar contamination. That means choosing a teacher model to generate data is already a choice of stance, and that step usually happens earlier than choosing a deployment model. The second thing is that evaluation methodology itself becomes a deployment variable. This report uses a single judge model, Grok 4.7, and a test set built by the party being evaluated; the definition of "propaganda" (repeating official narratives, endorsing them, or treating them as equally valid viewpoints) is itself positioned, and the report notes that the judge model is considered right-leaning politically and that language and answer format clearly affect results. So the percentages in the report are better treated as directional signals than as scores to put directly into a selection document. If you want to use such numbers in internal review, record which judge model was used and whether prompt language and format were standardized, or results from different batches won't be comparable.
Integration and evaluation
The following three suggestions are conditional and are not facts from the report. 1. Whether to adopt Arete-27B's weights directly or retrain with your own stance. The corresponding reported change is that about fifty million tokens of supervised fine-tuning can push one kind of answer tendency from 22% to below 1%, with general capability said to be largely intact. The choice depends on whether you want "neutral" or "a different explicit stance" — per the report, Arete-27B commits to Western values, whereas Nous's 2024 Hermes 3 took a neutral route configurable via system prompts. If your business requires the model not to take positions proactively (cross-border customer service, academic search), swapping in weights with an explicit stance just changes the direction of bias and still needs prompt-level constraints. A measurable acceptance criterion: take 100 to 200 sensitive-boundary prompts from your own business, each paired with 2 to 4 sourced reference facts, and use a fixed judge to measure, for the base model and candidate weights separately, the rate of contradiction with reference facts, the refusal rate and the deflection rate, rather than just an overall score. 2. Whether to build baselines by category or just compare overall model scores. The corresponding reported change is that propaganda rates differ noticeably across Chinese open-weights models (Qwen3.8-27B 22%, GLM-5.3 15%, DeepSeek V4.1 Flash 16%), and that on Arete-27B the biggest drops were concentrated in the free-speech category. The choice depends on whether your product will actually ask these questions: internal enterprise knowledge-base Q&A mostly won't, while retrieval-augmented generation or content generation on public issues will. A measurable standard is at least 20 prompts per category you actually cover, recording refusal rate, deflection rate and contradiction-with-reference-facts rate separately, and ranking by category rather than by overall model score; consider requiring contradiction rates in the categories you really trigger to fall below 5% before moving to the next selection step. 3. Whether to add topic-based spot checks to your synthetic-data pipeline. The corresponding reported change is that contamination may come from training data rather than the deployed model itself. The trigger is that you are using open-weights models to generate SFT data or judging data. A measurable standard: sample 50 to 100 generated instruction items per category across the seven categories, measure the contradiction-with-reference-facts rate using the same judging process as production, and make this check mandatory every time you change teacher models.
Limits and open questions
The percentages and scores in the report all come from Nous Research's own tests and self-reported results; The Decoder is relaying them, and this site has not independently reproduced them. The judge is a single model, Grok 4.7, and the test set was built by the party being evaluated; the 803 prompts span seven categories and are limited in scale. The report itself also notes the judge model's political lean and that language and answer format clearly affect results. The basis for "general capability intact" is Nous's internal Hermes Bench (65.60% vs. 65.77%) and Terminal-Bench 2.1 (67% to 70%, with Nous acknowledging a different test setup); the 3.07 score with reasoning mode enabled is based on only two of five runs. Training data has a median length of about 866 tokens, with most samples under 1,000 tokens, a distribution quite different from real long-context business requests. The contamination inference (that Nemotron 3 Ultra's anomalies come from synthetic data) and the claim that training/test prompt overlap was removed are both Nous's own statements, lacking third-party verification. Arete-27B does not currently appear in KimiSeek's model catalog; the release channel in the report is weights and data from Nous Research on Hugging Face, and availability, license terms and future updates are governed by the publisher — this does not mean it can be called through this site. The base Qwen3.8 27B has a corresponding entry in this site's catalog, but the two are different products and cannot substitute for each other. The report says the weights, test dataset and training data are released under the Apache 2.0 license; this is a license-level statement, and downloadable weights alone do not establish the scope of the grant — check the license text in the publisher's repository before actual use.
Illustrations from the original report

