Nous Research fine-tunes Qwen3.8-27B and cuts answers echoing Chinese state rhetoric from 22% to 0.9%
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Open weights allow a model's political framing to be changed substantially with relatively little fine-tuning while preserving its general capabilities; for now, however, only the developer's own methodology confirms the result.
Nous Research released the weights of Arete-27B under Apache 2.0, a version of Alibaba Qwen3.8-27B fine-tuned on answers from the Inkling model. Training comprised two passes over 28,886 examples, about 52.8 million tokens; reinforcement learning, preference optimization and architecture changes were not used. On Nous's own set of 803 prompts about freedom of expression, ideology, economics, education, privacy, democracy and the environment, the share of answers in which the model reproduced or endorsed China's official position fell from 22% for base Qwen to 0.9% for Arete with reasoning disabled. The average score for conveying specified facts rose from 2.32 to 2.98 out of 5. Grok 4.7 evaluated answers without being told which model wrote them; the methodology, data and evaluation code are public, but the results have not undergone independent review. On the internal Hermes Bench, the result remained almost unchanged at 65.60%, versus 65.77% for the base model.
Sources 2