Nous Research has published Arete-27B, a fine-tune of Alibaba's Qwen3.8-27B designed to change how the model responds to politically sensitive questions. The release gives developers both the weights and a concrete case study in how post-training can shift a model's answers. The model was released November 6; this report was published November 8.
What was changed
Nous says it used supervised fine-tuning on text data generated with an Inkling teacher. Its card describes roughly 52.8 million training tokens over two epochs, with language-model weights trained and the vision encoder frozen. It says the architecture was not changed. That is a different claim from proving the resulting model is free of bias.
Read the evaluation carefully
On the lab's 803-prompt evaluation, Nous reports that its chosen measure of Chinese-state-aligned framing fell from 22% for the starting model to 0.9% for Arete, while its fact score rose from 2.32 to 2.98 out of five. Grok 4.7 served as a blind model judge under the authors' rubric. These are developer-reported results on a deliberately selected test, not independent verification or a universal political-neutrality score. Some cross-model comparisons also used different sampling temperatures.
The evaluation dataset and training account allow others to inspect the methodology. The model card lists Apache 2.0; check component and downstream terms before reuse. AiBrainWorX has not run the model or reproduced the scores.
Why this matters
Our view: an inspectable fine-tune is valuable precisely because researchers can challenge both the training choices and the evaluation. Users should test several perspectives, examine the source material themselves and measure performance on their own tasks rather than assuming a new alignment produces objective answers.
Original reporting; claims attributed to their sources. No local tests implied.
