Nous Research releases a Qwen model retuned to reduce Chinese state-aligned answers

Developers now have a downloadable version of Alibaba’s Qwen3.8-27B designed to answer politically sensitive questions differently. Nous Research has released Arete-27B, reporting sharply fewer answers that adopt Chinese state narratives on its own test. The release includes training data and evaluation materials, exposing both the intended political shift and the rules used to measure it.

Nous describes the goal as reinforcing “Western values,” not building a model without a worldview. The lab says it observed refusals, evasions and Chinese Communist Party-framed answers in the starting model. Its response was to change the model through additional training, rather than leave users with the political defaults it shipped with.

A targeted rewrite, not a new foundation

Arete starts from Qwen3.8-27B rather than a model trained from scratch. Nous used supervised fine-tuning—teaching an existing model through examples—on about 52.8 million tokens, the chunks of text models process. Training ran for two passes through the data, without reinforcement learning, preference optimization or changes to the model’s architecture.

Nous selected Inkling as its teacher model, supplying answers for the training process. It generated prompts across seven value categories, checked for overlap with the separate evaluation set and removed overlapping prompts before training. The published training-data visualization lists 28,886 examples. Nous updated the language-model weights but left the vision encoder unchanged because the training material was text-only.

The judge checks facts and political framing

The evaluation contains 803 prompts covering freedom of expression, state ideology, markets, education, privacy and surveillance, democracy, and the environment. Each prompt has two to four required facts drawn from a dated source article, plus language associated with the official Chinese state position. Sources include English-language statutes, news and research reports; no single source supplied more than 13 prompts.

Grok 4.7 acts as the automated judge. It receives the question, answer, required facts and a fixed scoring guide, but not the identity of the answering model. Nous developed the judge using a 25-prompt pilot graded by humans, revising the guide three times after disagreements. One change made factual substance count rather than requiring exact wording, dates or citations.

The two scores measure different things. Factual recall runs from zero to five, based on whether required facts are stated, missed or contradicted. A single contradicted fact makes that answer score zero. The propaganda measure flags answers that adopt or endorse the state’s account, including presenting it as equally valid. Merely quoting that account to identify it does not qualify; a positive flag requires an exact supporting phrase.

Those headline comparisons use five runs of all 803 prompts, with thinking disabled and temperature set to zero. Nous reports improvements across all seven categories. Freedom of expression showed the largest shift: its factual score rose from 0.89 to 2.44, while the share flagged as propaganda fell from 60% to 1%.

Capability checks and an inspectable release

Changing political answers did not produce a meaningful capability regression on the tests Nous ran, the lab says. On its internal Hermes Bench, Arete scored 65.60%, against 65.77% for the base model—a difference Nous describes as indistinguishable from noise. It also reported a higher Terminal-Bench-2.1 result, but declined to claim an improvement because test configurations differed.

The release makes three parts of the work available for inspection. The model repository is on Hugging Face under an Apache 2.0 license, while the announcement links the data used to train and assess it:

• Model weights: the downloadable Arete-27B model, rather than access limited to a hosted service.
• Evaluation materials: all 803 prompts, source articles, required facts, the judge prompt and scripts for scoring other models.
• Training data: the released examples behind the additional training, allowing inspection of what Nous sought to teach.

Nous draws a broader conclusion: that the political bias layer is shallow and can be changed with carefully chosen supervision. But it also explicitly stops short of claiming to have remedied every behavior tied to the original viewpoint. The released evaluation offers a specific way to examine that argument: required facts, defined framing rules and a judge whose instructions others can inspect.