The American artificial intelligence laboratory Nous Research conducted an additional training procedure for the popular AI model Alibaba Qwen3.8-27B in order to rid it of its “Chinese roots.” As a result, she began to give answers that the developers of the original model would not always have considered acceptable.
Image source: alibabagroup.com
The Alibaba Qwen3.8-27B model, which has undergone additional training, has received a new name Arete; The goal of the project is to “establish Western values in a model originally trained in China.” The idea of the project was determined by the consideration that “Chinese models with open scales are among the most powerful today, but their weights initially contain behavior that corresponds to the state ideology of China,” explained the American developers.
The threats associated with Chinese AI models are not always limited to promoting Beijing’s ideology: under certain conditions, some of them can generate vulnerable program code. Hirundo’s application of retraining techniques to the Alibaba Qwen3.6-35B-A3B model, in particular, led to a sharp reduction in the amount of content reflecting Beijing’s preferences. If before the start of the process 89.8% of the model’s responses corresponded to these settings, then after retraining this figure dropped to 2.8%.
Nous Research found less impressive, although notable, results: the actual accuracy score for the Arete-27B model increased from 2.32 to 2.98 on a five-point scale, and the proportion of responses containing propaganda decreased from 22% to 1%. Such indicators are demonstrated by Western models tested in the laboratory. Nous Research looked at Nvidia’s Nemotron 3 Ultra when selecting Arete’s retraining tools, but found that it also “demonstrated a mixed approach to China-related topics.” They attributed this effect to the fact that when training Nemotron itself, Chinese models may also have been used.
As a result, the choice fell on the open model Inkling, released by the American laboratory Thinking Machines Lab. As experience has shown, the propaganda layer in Alibaba Qwen3.8-27B is superficial: it is relatively easy to eliminate using a training data volume of 50 million tokens.
If you notice an error, select it with the mouse and press CTRL+ENTER.
