Nous Research Re-trains Alibaba's Qwen Model with Western Values
US lab modifies Chinese AI to remove state-aligned censorship and propaganda.
1Nous Research Rebrands Qwen
Austin-based Nous Research has re-trained Alibaba’s Qwen3.8-27B model to create a new version named Arete. This process aims to remove state-aligned behaviors inherent in the original Chinese weights. The lab claims this re-education reinforces Western values, although the specific definition of these values remains open to subjective interpretation.
2Censorship in Chinese Models
Testing the Qwen3.8-27B-MLX-6bit model locally revealed significant limitations regarding sensitive historical topics. When queried about the 1989 Tiananmen Square events, the model either refused to provide information or offered euphemistic responses. This behavior highlights the underlying censorship baked into models originating from Chinese development environments and labs.
3Risks of Model Distillation
Anthropic has alleged that operators linked to Alibaba used outputs from its Claude models to train competing AI systems via distillation. Nous Research, developers of the Hermes Agent, views this practice as a security risk. They argue that using Chinese-trained models introduces potential vulnerabilities and unwanted propaganda into Western AI ecosystems.
4Vulnerability in Generated Code
Beyond ideological bias, concerns exist regarding the technical integrity of Chinese-made AI models. A study by Booz Allen Hamilton suggests that some of these models produce more vulnerable code under specific conditions. This technical risk adds another layer of scrutiny to the broader debate over Chinese influence in global technology.
5Machine Unlearning Success Rates
The vendor Hirundo reported success using machine unlearning on the Qwen3.6-35B-A3B model. Their process reduced outputs reflecting Communist Party of China propaganda from 89.8 percent down to 2.8 percent. This demonstrates that specific ideological layers within large language models can be systematically identified and significantly mitigated through targeted post-training interventions.
6Arete Model Performance Metrics
Nous Research reported that their Arete-27B model improved its fact score from 2.32 to 2.98 out of 5. Simultaneously, the lab successfully cut answers containing propaganda from 22 percent to 1 percent. These results align the model's output more closely with the performance standards observed in Western-developed AI models.
7Contamination of Training Data
Nous Research briefly considered Nvidia’s Nemotron 3 Ultra for the unlearning process but rejected it due to suspected bias. They theorize that Nemotron became contaminated by incorporating Chinese open models during its training. This highlights the difficulty of finding clean, high-quality base models in an ecosystem dominated by Chinese open-weight software.
8Shallow Nature of Propaganda
The lab ultimately utilized Inkling, an open-weight model from Thinking Machines Lab, to facilitate the re-education. Nous Research concludes that the propaganda layer in these models is surprisingly shallow. They demonstrated that with only 50 million tokens of supervised learning, this ideological conditioning is relatively easy to remove from the software.