Nous Research Releases Arete-27B, a Qwen3.8-27B Retrained to Drop the Chinese Official Line
Nous Research has open-sourced the weights for Arete-27B, a Qwen3.8-27B fine-tuned on roughly 53 million tokens to reduce answers aligned with Chinese state discourse. The lab also released its 803-question evaluation benchmark and training dataset.
Nous Research released Arete-27B on November 6, 2026, a retrained version of Alibaba’s open-weight Qwen3.8-27B model. Its stated goal: reducing refusals, evasive answers, and responses mimicking official Chinese discourse that the lab reports observing on politically sensitive topics. The weights are available on Hugging Face under an Apache 2.0 license, alongside the training data and evaluation suite.
What Nous Measured
Before training, Nous built a benchmark of 803 questions split across seven categories (freedom of expression, state ideology, state and market, education, privacy and surveillance, democracy, and environment). Each question is tied to a dated source article from which two to four expected facts were extracted, along with the Party’s official phrasing on the topic.
An automated judge (Grok 4.7) checks whether each response provides, omits, or contradicts the facts, and is required to cite verbatim any sentence adopting the official line as its own. Quoting this line to identify it (“the Party describes it as…”) does not count. The judge was calibrated on 25 hand-graded questions, undergoing three rubric revisions.
On this benchmark, Qwen3.8-27B echoes the official line in 22% of its answers, compared to 15% for GLM 5.3 and 16% for DeepSeek V4.1 Flash. The Western models tested (Mistral Large 4, Gemma 4 31B, Nemotron 3 Ultra, Muse Glimmer 30B) and Inkling all stayed below 1%.
How the Model Was Modified
The approach was intentionally straightforward: a supervised fine-tune of roughly 52.8 million tokens, two passes over the data, no reinforcement learning or preference optimization steps, and no architectural changes. The exemplar responses came from Inkling, the open model by Thinking Machines. Nous explained that it considered Nemotron 3 Ultra before discarding it because it also produced, according to its measurements, problematic answers regarding China; the lab hypothesizes contamination from data generated by Chinese models, without being able to prove it.
Around 20,000 questions were generated for training, then filtered to remove any overlap with the evaluation set. The research post reports 28,886 training examples in the published dataset.
The Reported Results
According to Nous, Arete-27B drops to around 1% of answers aligned with the official line, while its factuality score increases from 2.32 to 2.98 out of 5. The largest shift occurred in freedom of expression (dropping from 60% to 1% of aligned answers). In terms of general capabilities, the lab reports 65.60% on its Hermes Bench versus 65.77% for the base model, and 70% versus 67% on Terminal-Bench-2.1, measured using its own setup. These figures come directly from Nous, evaluated on a benchmark it designed and scored using another model.
Our take
The primary significance of Arete-27B lies less in the model itself than in the open methodology: an open benchmark, open data, and a low fine-tuning cost at the scale of a 27B model. Above all, it demonstrates that political behavior embedded in open weights can be altered with minimal resources, in either direction. The “Western values” framing is Nous’s own, and both the judge and the teacher are themselves models carrying their own baked-in choices. Before viewing this as definitive proof, we must wait for other teams to replicate the evaluation, including on topics unrelated to China. For production deployments, the practical question is not “which side” a model takes, but “which behavior, measured how”: and that is precisely what this benchmark enables teams to verify on their own models.
Sources
- Nous Research — Introducing Arete-27B (research post) Primary source
- Hugging Face — NousResearch/Arete-27B (weights, Apache 2.0 license) Primary source
- Hugging Face — NousResearch/arete-data-pairs (training data) Primary source
- Nous Research on X — Arete-27B announcement (November 6, 2026) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
