AI bias reduction in open-source models expands equitable access

Nous Research reduced political bias in Alibaba's Qwen3.8-27B model by 99.1% through targeted fine-tuning, releasing it under free Apache 2.0 licensing.

Nous Research fine-tuned Alibaba's Qwen3.8-27B model using 53 million tokens of supervised learning. Testing across seven categories—including free speech, surveillance, democracy, and environmental topics—showed the model's overall propaganda rate dropped from 22% to 0.9%. Factual accuracy improved from 2.32 to 2.98, while general capabilities remained stable at 65.60% on the Hermes Bench. The model was judged for compliance using xAI's Grok 4.7, with propaganda defined as responses reproducing Beijing's official narrative without factual correction. The team released Arete-27B under Apache 2.0 on Hugging Face, enabling open access.

This reduction in political bias directly addresses a key barrier to equitable AI access. By stripping away state propaganda from major models, the work lowers friction for users globally seeking unbiased knowledge tools and security applications. The open licensing ensures communities can adapt the model without gatekeeping.

What to watch: The Terminal-Bench 2.1 score improvement (67% to 70%) may reflect test setup changes rather than true capability gains. Additionally, training data examples were short (under 1,000 tokens), which could limit real-world robustness. The model's effectiveness for complex, nuanced political contexts remains unverified.

Source: The Decoder