Arete-27B shows how sharply post-training can rewrite model values

Nous Research started from Alibaba's Qwen3.8-27B and trained Arete-27B with roughly 50,000 examples, 52.8 million tokens and 1,202 supervised fine-tuning steps, releasing the 27B weights, training pairs and an 803-prompt evaluation. With reasoning off, the developer reports that answers flagged by its rubric for propaganda framing fell from 22% to 0.9% and the fact score rose from 2.32 to 2.98; Hermes Bench stayed roughly flat and Terminal-Bench moved from 67% to 70%. These figures use Nous's own “Western values” categories and a Grok 4.7 judge, not an independent neutrality certificate. The stronger finding is that political behavior in a capable model can be changed quickly with relatively modest post-training—and could be changed in the opposite direction too.