Nous Research has retrained Alibaba's large language model Qwen3.8-27B to create a new version called Arete-27B. This retraining significantly lowered the model's pro-Chinese propaganda bias, cutting it from 22 percent to under one percent while preserving its overall capabilities.
The reduction was achieved through supervised fine-tuning on 53 million tokens, according to The Decoder. This approach demonstrates how targeted retraining can improve the neutrality of AI models without sacrificing performance.
For Japanese investors and tech firms, this development highlights growing efforts to address geopolitical bias in AI, which is increasingly important as such models are integrated into global financial and technology markets.