What changed
Nous Research announced it fine-tuned Alibaba's Qwen3.8-27B to remove Chinese state propaganda from the model's responses. The resulting model, Arete-27B, reduces the propaganda rate from 22 percent to under one percent while maintaining general capabilities. The fine-tuning process used 53 million tokens of supervised training data over two epochs, without reinforcement learning or preference optimization.
What this means in practice
The fine-tuned model, Arete-27B, is available under an Apache 2.0 license on Hugging Face. It scores 2.98 on factual accuracy and shows a significant improvement in the free speech category, with the fact score jumping from 0.89 to 2.44 and the propaganda rate falling from 60 to one percent. The model's general capabilities remain largely unaffected, with scores on internal benchmarks like Hermes Bench and Terminal-Bench 2.1 showing minimal differences compared to the base model.
Limitations and availability
The training data for Arete-27B is spread nearly evenly across seven value categories, with most examples under 1,000 tokens. The model's propaganda rate is on par with Western models, while other Chinese models like GLM-5.3 and DeepSeek V4.1 Flash score much higher. The weights, test dataset, and training data are publicly available, but the model itself is not yet widely adopted. The fine-tuning process did not involve reinforcement learning or preference optimization, and the team removed overlaps between training and test prompts beforehand.
