American artificial intelligence lab Nous Research has completed a fine-tuning procedure on Alibaba's popular open-weights model Qwen3.8-27B to, in the developers' words, "strip away its Chinese roots." As a result of the retraining, the model was given a new name—Arete—and began generating responses that the original model's creators at Alibaba likely wouldn't consider acceptable. The stated goal of the project is formulated directly: "to instill Western values in a model originally trained in China."
Why "Chinese roots" became a problem
The project's intent stems from the view that Chinese open-weights models are among the most powerful available today, yet their weights, according to American developers, "inherently embed behavior aligned with China's state ideology." As Nous Research explains, this makes the direct use of such models risky for Western products and services that require neutral outputs or ones compliant with local standards.
Threats beyond ideology: Vulnerable code
Risks associated with Chinese AI models, according to the researchers, are not always limited to promoting Beijing's ideology. Under certain conditions, some of these models are capable of generating vulnerable program code, posing a practical threat to application security. This broadens the motivation for "cleansing" open weights of behavioral patterns baked in during training.
The Hirundo precedent: From 89.8% to 2.8%
A telling example of the effectiveness of retraining was demonstrated by Hirundo, which applied fine-tuning methods to Alibaba's Qwen3.6-35B-A3B model. As a result, the volume of content reflecting Beijing's preferences plummeted: while 89.8% of the model's responses aligned with these guidelines before the process began, the figure dropped to 2.8% after retraining. This case confirms that the ideological layer in open Chinese models is amenable to correction.
How Arete was retrained
Nous Research researchers achieved less dramatic but noticeable results: Arete-27B's factual accuracy rose from 2.32 to 2.98 on a five-point scale, and the share of propaganda-containing responses dropped from 22% to 1%—figures comparable to Western models tested in the lab. When choosing a tool for retraining, specialists considered Nvidia's Nemotron 3 Ultra model, but found that it too "showed an ambiguous approach to China-related topics," which was attributed to the potential use of Chinese models in training Nemotron itself. Ultimately, the choice fell on Inkling, an open model released by the American lab Thinking Machines Lab. Practice showed that the ideological layer in Alibaba Qwen3.8-27B is superficial and relatively easy to eliminate using a training volume of about 50 million tokens.
What this means for the open-weights market
The Arete case illustrates a growing trend of "re-tuning" powerful open Chinese models to match Western behavioral and safety standards. On the one hand, this confirms that the ideological component in such models' weights is not irreversible and can be removed with a relatively small amount of data. On the other hand, it raises the question of how deeply and sustainably one can control the behavior of open weights, which, once published, circulate freely and can be modified by any user.
