Nous Research has released Arete-27B, a 27-billion-parameter fine-tune of Alibaba’s Qwen3.8-27B designed to change answers that adopt Chinese-state framing. The lab describes its aim as reinforcing “Western values”, making the project an explicit attempt to alter a model’s viewpoint rather than a claim to have built a politically neutral assistant.
For people choosing open-weight models for research or agentic work, the release offers weights and an evaluation that can be inspected. The headline result is narrower than a general finding about bias: Nous says the share of answers flagged as containing propaganda fell from 22% to 0.9% on its own test. All the reported evaluation results are Nous’s measurements, not independent validation.
Arete-27B’s 0.9% result uses Nous’s own definition
The Hugging Face model card describes an evaluation of 803 prompts across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. Each prompt comes with two to four required facts taken from a dated source article. The source material includes English-language statutes, news and research reports.
Nous used Grok 4.7 as an automated judge, supplying the answer, required facts and a fixed grading rubric without identifying which model produced the response. The lab says it developed the judge using a 25-prompt, hand-graded pilot and revised the rubric where human and automated grades disagreed.
On this assessment, the average fact score rose from 2.32 to 2.98 out of five. That scale is not a percentage of universally correct statements: an unanswered prompt scores zero, as does an answer opposing any required fact. The card also reports that answers scoring zero fell from 359 to 216 out of 803 in the comparison.
The separate propaganda measure flags answers that adopt the Chinese state’s account as their own, either replacing a required fact with an official line or endorsing that line while stating the facts. Merely quoting the official position to identify it does not qualify under the rubric. The reported 0.9% is the share of answers receiving either of the two non-zero framing scores, not proof that all ideological content or censorship has disappeared.
Each main comparison row covers five runs. Arete and the starting Qwen model were tested with thinking disabled at temperature zero, while other models in the wider comparison table used temperature 0.6. The starting-model comparison therefore shares those settings, but the table as a whole is not a uniform cross-model test.
Supervised training preserves architecture, with mixed benchmark movements
Nous says it generated 20,000 prompts across the seven categories, removed overlap with the held-out evaluation and collected responses from a teacher called Inkling alongside the starting model’s answers. It trained on about 52.8 million tokens at 64k context over two epochs and 1,202 steps. The process used supervised fine-tuning: learning from example responses, without reinforcement learning, preference optimisation or architectural modification.
The repository’s model-card metadata lists the Apache 2.0 licence. Unlike the earlier MISSINGNO offering through Nous Portal, whose announcement left important model details unstated, Arete comes with published weights and a detailed account of its training and assessment. Nous distinguishes the released training dataset from its separate held-out evaluation dataset.
Nous’s internal Hermes Bench result moved from 65.77% for Qwen3.8-27B to 65.60% for Arete. The lab interprets that small difference as noise. Terminal-Bench-2.1 scores were 67% and 70%, respectively, but Nous expressly declines to take credit for the apparent improvement, noting that inference configurations and harnesses can produce differences from official benchmark results. It says its own starting-model and fine-tuned runs used matching local configurations.
Those results address selected general and agentic tasks, not every possible consequence of changing a model’s behaviour. Another adaptation of the same starting model, Underdog’s Saluki 27B, reported tool-calling gains alongside a mathematics decline in its own tests, illustrating why separate capability measurements matter.
Arete also sits alongside Nous’s broader push to build Hermes for businesses. Here the release is a model and its supporting research material, rather than the agent harness or an enterprise deployment announcement.
Less flagged framing is not the same as neutrality
The most useful distinction is between changing a specified behaviour and removing bias in general. Nous has chosen the topics, reference facts and definition of framing it wants to measure. Publishing those choices makes the result more inspectable, but a lower score still means closer alignment with that particular rubric. It does not establish agreement across other political traditions, languages or contested questions outside the test.
The judge’s blind treatment of model identity addresses one possible source of favouritism without making the evaluation independent of its designer. Likewise, matching decoding settings for the starting model and fine-tune strengthens that specific comparison; it does not erase the different settings elsewhere in the table. Readers can preserve the useful within-pair result without treating every row as a league table on equal terms.
For deployment decisions, the framing assessment and capability tests answer different questions. An assistant can change the way it discusses political subjects while retaining similar scores on selected tasks, yet still require testing against the actual work and failure modes a user cares about. The next consequential evidence would be independently reproduced results and tests beyond Nous’s chosen categories, using the published evaluation material rather than treating the release’s stated values as a guarantee.
Sources
Related stories
Underdog’s Saluki 27B reports tool-calling gains and a maths decline is a useful next read on a different adaptation of Qwen3.8-27B and the need to examine task-specific trade-offs.
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
