State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away

Why it matters

Nous Research has released Arete-27B, an open-weight model created by fine-tuning Alibaba's Qwen3.8-27B to remove Chinese state-mandated behavior from its responses. The US startup, known for its open-source agent Hermes, observed that the base model produced refusals, evasive answers, and Chinese Communist Party-aligned responses, and it used supervised fine-tuning on a large token dataset to strip that layer away. Nous measured the change with its own benchmark covering categories such as free speech, state ideology, and surveillance, finding that propaganda nearly disappeared while general capability scores on its internal Hermes Bench stayed essentially flat. Nous argues the result shows political bias in Chinese open-weight models is a thin, late-applied layer rather than something deeply embedded in the weights. The move also marks a strategic shift: unlike Hermes, which was deliberately neutral, Arete is explicitly aimed at what Nous calls Western values. The weights, test dataset, and training data are public, allowing others to verify the claims.