Nous Research has released Arete-27B, a fine-tuned version of Alibaba's Qwen3.8-27B designed to change how it answers politically sensitive questions. The lab's November 2026 announcement describes an effort to reinforce Western values. Developers can inspect the weights and evaluation, though the reported improvements come from Nous's own tests. MrKeyoor has not rerun them.
The release puts a specific question within reach of independent testing: how much can post-training change a model's political responses while preserving its usefulness elsewhere? Nous says it used supervised fine-tuning, without reinforcement learning or architectural changes, according to the model card.
Two scores that need to stay separate
Nous reports that Arete's factual-recall score rose from 2.32 to 2.98 out of five against its starting model. Answers flagged for adopting the Chinese state's account fell from 22% to 0.9% in the published model-card comparison. Both models ran with thinking disabled at temperature zero.
The released evaluation explains why those numbers belong together. It scores factual content and political framing independently. An unanswered question receives zero for both. A model that stops answering could therefore reduce its propaganda rate while becoming less useful. Arete's higher fact score matters because it checks that particular failure mode.
The test contains 803 prompts. Nous's announcement says Grok 4.7 graded responses against required facts and a fixed rubric, without knowing which model produced each answer. The lab developed the judge using a hand-graded pilot of 25 prompts. Those choices make the procedure inspectable, but the scores still depend on the selected questions and judge.
What a developer can check next
The training used Inkling as a teacher and about 52.8 million tokens. Nous's model card lists the weights under Apache 2.0. This gives researchers a concrete derivative to examine rather than just a claim that a different viewpoint is possible.
Nous also reports little change on its internal Hermes Bench, scoring 65.60% against Qwen's 65.77%, in the announcement. That supports a narrow capability-retention claim on the lab's tests. It cannot establish that every application will behave the same way.
The next useful evidence is an independent rerun with matched sampling settings, followed by questions outside this test. The evaluation's limitations specify English-only coverage and warn that different judges produce different absolute scores. Until broader tests arrive, Arete offers a way to investigate adjustable model behavior, with no basis here for calling the result politically neutral.