State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away

Fine-tuning can remove state propaganda from Chinese AI models, making them usable for Western markets.

Nous Research, a US-based AI startup, has developed a new open-weight model named Arete-27B by fine-tuning Alibaba's Qwen3.8-27B model. The process involved training on 53 million tokens over two epochs, which significantly reduced the model's Chinese state propaganda output from 22 percent to under one percent. This reduction was achieved without compromising the model's general capabilities, as evidenced by performance on Nous's internal Hermes Bench and Terminal-Bench 2.1 tests. Nous developed a specific test suite with 803 prompts across seven categories to measure propaganda, using Elon Musk's xAI Grok 4.7 as a judge model. The resulting Arete-27B model is positioned to favor "Western values" over Beijing's official narrative, making it comparable to Western AI models in terms of propaganda rate.

The fine-tuning process demonstrates that politically biased behavior in Chinese open-weight models is superficial and can be effectively removed with targeted data. This finding is critical as it suggests that organizations can leverage the strong foundational capabilities of Chinese models without inheriting their embedded state-aligned messaging. By releasing the model and training data, Nous Research is enabling broader scrutiny and adoption of less biased AI, potentially influencing the development and perception of global AI safety standards.

This development suggests a potential shift in the open-source AI landscape, where models originally trained with specific national biases can be repurposed to align with different geopolitical or ethical frameworks. It impacts users and developers who require models free from state propaganda for applications operating in Western markets or adhering to Western values. The ability to strip away this bias quickly and effectively means that the foundational strengths of models like Qwen can be utilized without the political baggage, opening new avenues for collaboration and competition.