Nous Research says its Arete-27B model adopted the Chinese state’s account in 0.9% of answers in a defined test, compared with 22% for Qwen3.8-27B. That is a striking result for one specific kind of political framing, not a general certificate of neutrality.
Nous/Hermes Watch analysis
What happened
Nous introduced Arete-27B on 6 November as a supervised fine-tune of Qwen3.8-27B. In the evaluation described by NeoTeo, Nous used 803 prompts across seven categories, including democracy, surveillance, education and freedom of expression. Each model was tested five times, with reasoning off, temperature set to zero and a 32,000-token limit.
Nous counted an answer as adopting the Chinese state’s account when it met a defined scoring threshold. Grok 4.7 judged the answers without being told which model produced them. Nous reported a 21.1 percentage-point difference between the models, using its rounded rates.
Why it matters
The result gives a concrete example of fine-tuning aimed at a particular model behaviour, with a prompt set and scoring rules readers can inspect. It also shows why the label “political bias” needs more care than a single headline can provide: the result is about a defined set of prompts, a specific framing and Nous’s scoring method, not every political question a model might answer.
Nous says the work used supervised fine-tuning with the open teacher model Inkling, without reinforcement learning or preference optimisation. It also describes the difference on Hermes Bench as noise and does not claim a Terminal-Bench improvement.
Our read
A 0.9% figure is attention-grabbing, but the useful story is the test behind it. Arete-27B comes with a defined evaluation and, according to the report, links to its weights, training data and prompts under an Apache 2.0 licence. That makes the claim more inspectable than a vague promise of “neutrality”. It does not make the test universal, and a tidy percentage is not a personality transplant.
What to watch
- Whether independent researchers reproduce the result using the published prompts and scoring rubric.
- How Arete-27B performs on other political framing tests and in ordinary use.
- Whether Nous publishes further results that clarify the trade-offs from fine-tuning.
Discussion spark: Is a narrow, reproducible political-framing test more useful than a broad claim of model neutrality, or does it risk making one particular rubric the definition of fairness?
Sources and evidence
- Qwen 3.8 pro-Beijing answers in Nous's Arete-27B test – NeoTeo (10 November 2026, 12:58 UTC)
Nous/Hermes Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Nous Research.