US start-up Nous Research has fine-tuned Alibaba's open-weight model Qwen3.8-27B to remove Chinese state propaganda from its answers. The new model is called Arete-27B. According to Nous, the share of propagandistic answers in its own test fell from 22 to 0.9 percent. These figures come from Nous itself and have not yet been independently verified.
Method
Nous trained Qwen3.8-27B using supervised fine-tuning on roughly 53 million tokens over two epochs, without reinforcement learning or preference optimization. Inkling from Thinking Machines Lab served as the teacher model, because an open model was meant to avoid trouble with terms of use that prohibit distillation. Nvidia's Nemotron 3 Ultra was also considered but, according to Nous, showed irregularities on China-related topics.
The test
Nous built a test with 803 prompts in seven categories, including freedom of expression, state ideology, data protection, surveillance and democracy. For each prompt the team defined two to four facts from a dated source article. Grok 4.7 from SpaceXAI served as the judge model, checking whether an answer states, omits or contradicts these facts. The unmodified Qwen3.8-27B scored 2.32 out of 5 on facts, and 22 percent of its answers contained propaganda. Arete-27B reaches 2.98 points. On freedom of expression, the propaganda rate fell from 60 to one percent.
General capabilities
According to Nous, performance does not suffer: on its internal Hermes Bench, Arete-27B scores 65.60 percent versus 65.77 percent for the base model. On Terminal-Bench 2.1 the score rises from 67 to 70 percent, but Nous does not draw a true improvement from this because of differing test configurations.
Background
The political slant of Chinese models is documented several times over. NewsGuard found in 2025 that five leading models, including Qwen, repeated or failed to correct pro-Chinese false claims in 60 percent of cases. A study by Aleph Alpha with 967 politically sensitive questions found only 17 to 41 percent balanced answers from models by Alibaba, Deepseek and Moonshot AI.
Assessment
If the results hold up, the project suggests that political defaults in open models can be changed with relatively little additional training. For companies using Chinese open-weight models, this raises questions about the neutrality and verifiability of model outputs. It remains open whether the measurement, based on a US judge model and an in-house test set, is independent enough. Weights, test dataset and training data are available on Hugging Face under the Apache 2.0 license.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
