Nous Research, on November 6, released Arete-27B, a 27-billion-parameter open-weight model built via supervised fine-tuning of Alibaba’s Qwen3.8-27B. The company said the model reduces what it calls propaganda-style framing from 22% to about 0.9% while holding general capability roughly level on its internal benchmarks.
The training recipe was deliberately minimal. Nous used about 52.8 million tokens of teacher responses from Inkling, an open model from Thinking Machines Lab, over two epochs and 1,202 steps. There was no reinforcement learning, no preference optimization, and no architectural change. The company published the weights, training pairs, and an 803-prompt evaluation set on Hugging Face under an Apache 2.0 license.
The lab did not present the model as neutral and said:
“We post-trained over Alibaba’s Qwen3.8-27B to reinforce Western values on a Chinese-trained model,” Nous wrote in the release.
The name Arete comes from the ancient Greek word for excellence or virtue.
The headline numbers come from Nous’s own evaluation, which covers seven categories:
- freedom of expression,
- state ideology,
- state versus free market,
- education,
- data privacy and surveillance,
- democracy,
- and environmental policy.
Each prompt carries two to four required facts drawn from dated source articles. An LLM judge, xAI’s Grok 4.7, scored whether answers stated, missed, or contradicted those facts and whether they adopted the official Chinese line.
Nous said it chose the prompts from English-language government statutes, news reports, and research documents, with no single source supplying more than 13 prompts.
On that test, Arete-27B’s factual recall score rose from 2.32 to 2.98 out of 5. The propaganda rate fell from 22% to 0.9%. The largest shift was in freedom of expression, where the rate dropped from 60% to 1%.
Nous also reported that the share of answers that scored zero on facts fell from 359 to 216 across 803 prompts.
What the propaganda evaluation found
The central part of the release is an evaluation developed by Nous Research containing 803 prompts across seven value categories.
Each prompt included between two and four required facts extracted from dated source material, alongside language representing the Chinese government’s official framing of the subject.
The researchers measured two things: whether a model recalled the required facts and whether it adopted the Chinese state’s position as its own.
Nous used Grok 4.7, developed by xAI, as an independent model judge for the exercise. The judge received the questions, model responses and scoring criteria without being told which model had generated each answer.
According to the published results, Arete-27B substantially reduced the frequency of answers classified as propaganda.
The comparison included several other models:
| Model | Factual recall (out of 5) | Answers containing propaganda |
|---|---|---|
| Inkling | 3.96 | 0.3% |
| Mistral Large 4 | 3.65 | 0.3% |
| Muse Glimmer 30B | 3.02 | 0.5% |
| Gemma 4 31B | 2.58 | 0.5% |
| Arete-27B | 2.98 | 0.9% |
| GLM 5.3 | 3.10 | 15% |
| DeepSeek V4.1 Flash | 2.39 | 16% |
| Qwen3.8-27B | 2.32 | 22% |
The largest improvement appeared in the freedom-of-expression category.
There, the original Qwen model recorded a factual recall score of 0.89, while Arete-27B achieved 2.44. The proportion of responses classified as propaganda fell from 60% to 1%.
The number of answers receiving a zero factual score also decreased, from 359 to 216 across the 803 prompts.
Nous said the improvements appeared across all seven categories.
The company also published the complete evaluation prompts, source articles, scoring instructions and scripts. This makes it possible for researchers to reproduce the tests rather than relying exclusively on the headline figures.
Did the model lose its general capabilities?
Changing a model’s political behaviour through fine-tuning can affect other abilities, including coding, reasoning and general instruction-following.
Nous tested Arete-27B against its original Qwen model using two benchmarks.
On Hermes Bench, the company’s internal general-capability evaluation, Arete scored 65.60%, compared with 65.77% for the original Qwen3.8-27B.
The difference was less than 0.2 percentage points.
On Terminal-Bench-2.1, which evaluates an AI model’s ability to complete tasks in a terminal environment, Arete scored 70%, compared with 67% for the original model.
Nous did not attribute the improvement in Terminal-Bench performance to its fine-tuning, citing differences between benchmark configurations and evaluation environments.
The results suggest that the additional training did not produce a substantial decline in the capabilities measured by these two tests.
General capability stayed close to the base model.
The wider debate over political bias in AI
Arete-27B arrives as researchers increasingly examine how political environments, training material and model-development decisions influence AI-generated answers.
A separate study published by AI evaluation company Vals AI in October 2026 found that political positioning in Chinese and Western AI models is more complicated than a simple division between Chinese and American systems.
Its research tested models using hundreds of politically sensitive questions, different languages and different deployment contexts.
Vals AI found that Chinese models sometimes criticised their own government, while some non-Chinese models occasionally repeated Chinese Communist Party narratives.
It also found that the language used to question a model could change its answers. Chinese models generally expressed more favourable positions towards China when prompted in Simplified Chinese than when asked comparable questions in English.
The study further found that placing models inside agent frameworks could reduce certain forms of political alignment.
These findings provide additional context for Nous Research’s work. Political framing is not exclusive to one country’s AI systems, and the way models are trained and evaluated can influence the answers they produce.
Arete-27B takes that issue into the model-training process itself. Instead of relying on system prompts to instruct an AI to avoid particular political positions, Nous changed the underlying model weights through additional training.
Open weights and training data released
Nous Research has made Arete-27B available through its Hugging Face repository under the Apache 2.0 licence.
The release includes the 27-billion-parameter model weights, its model documentation and links to the associated training and evaluation datasets.
The company says the project demonstrates how relatively small amounts of additional training can change the political framing of an existing open-weight model without requiring a new architecture or a complete training process.
Nous Research wrote in its announcement:
“The values in any open model are only ever one fine-tune away.”
Hermes Bench, Nous’s internal suite of 150 agentic tasks, came in at 65.60% for Arete-27B against 65.77% for Qwen3.8-27B. Terminal-Bench-2.1 scored 70% for Arete-27B against 67% for the base model, though Nous attributed possible differences to inference configuration rather than training.
The release lands in a policy environment where US agencies have restricted Chinese AI models from federal use, and Nous framed the work as a way for companies and agencies blocked from Chinese open models to run a capable 27B model locally.
Nous argues that state-aligned behavior in Chinese open-weight models is a thin layer applied late in training, and that a relatively small fine-tune can shift it without wrecking capability.
Outside researchers have not replicated the results yet. The evaluation was designed by Nous, the judge model was chosen by Nous, and the prompts were written by Nous.
The training-data card also warns that most accepted teacher examples still carry a factual_error tag, which means the fine-tune was partly trained on answers Nous’s own pipeline flagged as containing factual mistakes.
Nous said it published the evaluation set so others can rerun it, but as of November 9 no independent rerun had surfaced.
The company, founded in 2023 and based in Austin, Texas, raised $90 million in October at a reported $1.5 billion valuation. Its Hermes Agent is MIT-licensed and free to self-host. Arete-27B is available for local inference through Nous and Hugging Face.
