A new testing suite consisting of 803 prompts revealed that a recently modified large language model is far less likely to adopt the official party line of the Chinese government on issues like democracy and privacy. The frequency of state-aligned responses in the updated 27B-parameter assistant fell from 22% to 1% after it was post-trained on approximately 53 million tokens.
The new model, called Arete-27B, is a version of the Qwen3.8-27B open-weights system that Nous Research modified to prioritize Western values. The developer stated that the lauch of the new weights includes no performance penalty based on Hermes Bench scores, and has made the weights and training data available to the public.
