Nous Research Launches Arete-27B to Strip State Alignment From Qwen AI Model

Nous Research released Arete-27B, a modified version of the Qwen3.8-27B open-weights model post-trained on approximately 53 million tokens to reduce state-aligned responses, with a testing suite of 803 prompts showing the frequency of such responses falling from 22% to 1%. The developer stated that the launch includes no performance penalty based on Hermes Bench scores and has made the weights and training data publicly available.