Nous releases Arete-27B to let builders change a model’s political framing

Nous Research has released Arete-27B, a downloadable model designed to change how Alibaba’s Qwen3.8-27B handles political and value-laden questions. The repository history records the initial upload on November 6, and the model card lists an Apache 2.0 license.
The 27-billion-parameter model is a supervised fine-tune rather than a new architecture. Nous says it used text-only post-training and an Inkling teacher model, without reinforcement learning or preference optimization. Its stated purpose is to reinforce Western values and reduce answers that adopt Chinese-state framing.
A measurable change with a narrow definition
Nous’s launch report describes an evaluation of 803 prompts across seven categories, including expression, democracy, surveillance and environmental policy. It separates whether an answer supplies required facts from whether it endorses the Chinese government’s framing.
On that evaluation, the lab reports that the share of responses flagged for state framing fell from 22% for the starting model to 0.9% for Arete with thinking disabled. The average factual score rose from 2.32 to 2.98 out of five.
Those figures describe performance under Nous’s rubric, judged by another language model. They do not establish that Arete is politically neutral or more accurate across unrelated subjects. The training objective itself selects a viewpoint, and users may disagree with parts of that selection.
Nous has also linked the evaluation and training dataset from its announcement. Making those materials available gives other researchers a way to inspect what the lab trained toward and challenge how it measured the outcome.
A candidate for controlled local comparisons
The weight repository totals about 54.7 GB. That makes memory planning part of deployment: stored weights are only one component of a running model, alongside caches and the inference software.
For builders, the useful experiment is a comparison against the unmodified base on questions their application actually faces. A good test would separate factual mistakes, inappropriate refusals and changes in tone or interpretation, rather than combine them into one “bias” score.
It should also include routine coding and instruction-following tasks. A model that changes political framing still needs to preserve the behavior the application depends on.
Arete’s practical contribution is a downloadable, inspectable example of changing model behavior after pretraining. Whether its choices fit a particular product remains a deployment decision, best made with explicit evaluation criteria rather than a neutrality label.
Sources and citations
Revision history
Published .