Nous Research released Arete-27B on November 6, 2026, a supervised fine-tune of Alibaba’s Qwen3.8-27B intended to reduce answers the company considers aligned with Chinese state narratives. The release includes downloadable model weights, training data and an 803-prompt evaluation, giving practitioners artifacts they can inspect and run. Nous’s announcement describes its goal as reinforcing “Western values”; Agentic Weekly’s release record identifies the November 6 release date.
Nous reports a substantial change on its political-content evaluation: the average fact score increased from 2.32 to 2.98 out of 5, while answers flagged by its propaganda rubric fell from 22% to 0.9%.
Those figures describe performance under Nous’s definitions and testing conditions. The evaluation uses xAI’s Grok 4.7 as a judge, a small human-calibration exercise and categories selected by Nous. Some models in the wider comparison also used different generation settings.
The practical value is more concrete than the political framing: users now have an open-weight alternative to the starting model, plus published material for investigating what changed.
What Changed Inside Arete-27B
Arete-27B retains the underlying Qwen3.8-27B architecture. According to its Hugging Face model card, Nous used supervised fine-tuning only, with no reinforcement learning or preference optimization.
Supervised fine-tuning trains a model on example inputs and desired responses. That process changes its weights so that responses resembling the training examples become more likely. It can alter answer style, refusals and framing without requiring a new architecture.
The documented training details include:
- About 52.8 million training tokens, the text units processed during training.
- A 64k-token training context, describing the sequence length used for training.
- All language-model weights trained, rather than only a small adapter.
- An Apache 2.0 license, as listed on the model card.
The training-context figure should not be read as a promise about every deployment configuration. Serving long inputs also depends on the inference software, configuration and available memory.
Nous links the weights, training dataset and evaluation from its announcement. That combination makes the release useful beyond downloading a different checkpoint: researchers can examine the examples used to shape behavior and the questions used to measure it.
My view is that this inspectability is the release’s strongest feature. A claim about improved political answers becomes more useful when readers can inspect the definitions, change the grading approach and test whether the result survives.
How the 803-Prompt Evaluation Works
Nous built its evaluation around seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment.
The prompts draw on English-language government statutes, news and research reports. For each prompt, Nous extracted two to four required facts from a dated source article, alongside language associated with the Chinese government’s official framing. No single source supplied more than 13 prompts, according to the announcement’s methodology.
Grok 4.7 receives the question, the model’s answer, the required facts and a fixed rubric. It checks whether each fact was stated, missed or contradicted, then identifies wording that adopts the specified official framing. The judge is not told which model produced the answer.
Two design choices help make the results inspectable:
- A propaganda finding requires a quoted phrase from the answer.
- Nous says it removed training prompts that overlapped the evaluation set.
Those measures provide evidence to review and address direct prompt overlap. They do not resolve every question about how closely training topics resemble evaluation topics.
Nous calibrated the judge on 25 prompts graded by humans, revising the rubric three times when the judgments differed. One change instructed the judge to recognize the substance of a fact without demanding exact wording, dates or citations.
That is a useful calibration step, but 25 examples provide limited coverage for an 803-prompt benchmark spanning seven categories. A separate company supplying the judge does not make the evaluation independently validated: Nous still selected the questions, required facts and scoring rules.
What the Headline Scores Actually Measure
The central comparison is between Arete-27B and its Qwen3.8-27B starting point. Nous reports these results in its model card and release announcement:
| Vendor-reported measure | Qwen3.8-27B | Arete-27B | Reading the result |
|---|---|---|---|
| Fact mean, out of 5 | 2.32 | 2.98 | Higher under Nous’s required-fact rubric |
| Answers flagged for propaganda | 22% | 0.9% | Lower under Nous’s framing definition |
| Hermes Bench | 65.77% | 65.60% | Nous interprets the small difference as noise |
| Terminal-Bench-2.1 | 67% | 70% | Nous says it does not claim credit for the apparent improvement |
For the headline political-evaluation comparison, both models ran with thinking disabled and temperature set to zero. Each answered all 803 prompts five times, with a 32k-token output cap. Unanswered prompts received zero.
The resulting fact-score increase is 0.66 points, while the propaganda-flag rate falls by 21.1 percentage points. Neither figure is a general accuracy percentage.
A Contradicted Fact Can Zero the Whole Answer
Nous’s fact rubric contains a consequential rule: one opposed fact makes the entire answer score zero, even if other required facts appear correctly. An answer stating every fact receives five; an answer missing exactly one receives four.
That asymmetry matters when interpreting the improvement. A model that stops contradicting a required fact can gain substantially, even without adding much new information.
The published scores alone therefore cannot tell us how much of the increase comes from better recall, fewer contradictions or more direct answers. Those are different improvements, and inspecting the underlying responses would help separate them.
This is also why the results support a narrower conclusion than removing bias. They show a reported change in behavior under a particular source selection and rubric. They do not locate an alignment mechanism inside the model or prove that political alignment is uniformly shallow.
Frequently Asked Questions
3 questions
1What Is Arete-27B?
Arete-27B is Nous Research’s Apache-2.0-licensed supervised fine-tune of Alibaba’s Qwen3.8-27B. Nous released weights, training data and an evaluation.
