Announcement
Nous Research releases Arete-27B, a Qwen model tuned to reinforce Western values
Nous Research says Arete-27B cut answers flagged as Chinese-state propaganda from 22% to 0.9% in its evaluation.
TLDR
Built from Alibaba’s Qwen3.8-27B, Arete-27B received about 52.8 million tokens of additional training. Nous Research used an AI judge to assess factual recall and Chinese-state framing across 803 prompts in seven value categories. The company reports a fact score of 2.98 out of five, versus 2.32 for the base model. It released the model’s weights, training data and evaluation.
Combined views
—
2 Sources, first seen ago
