A US lab has taken Qwen3.8-27B, Alibaba's open-weight model, and retrained it with a post-training process that, in the company's own words, aims to «reinforce Western values» on a model trained in China. The result is called Arete-27B.

Psst!Add AlexTech.ai asPreferred Source on Google

The project comes from Nous Research, an Austin, Texas-based startup known for Hermes Agent, an open-source personal agent. The company describes the process as «re-education through training», explicitly distinguishing it from China's re-education through labour system. It also notes that what counts as a «Western value» remains open to interpretation.

official image of Qwen3.8-27B

Image: AlibabaCloud-Official — https://github.com/AlibabaCloud-Official/Qwen3.8-27B

The stated rationale is that open-weight Chinese models are among the strongest available, but «ship with state-aligned behavior baked into the weights». For this reason, the company argues, adopting them carries a risk for users.

The work used 52.8 million tokens of supervised fine-tuning over two epochs, with no reinforcement learning and no architectural changes. To assess the result, Nous Research built an 803-prompt test across seven value categories, using a judge it developed internally. According to the company's figures, the share of answers containing propaganda fell from 22% to 1%, while the fact score rose from 2.32 to 2.98 out of 5. These are internal measurements, not verified by third parties.

Nous Research also says the work shows the propaganda layer is shallow: with about 50 million tokens of supervised learning, it argues, it is relatively easy to remove.

A technical detail stands out. Nous Research first considered Nvidia's Nemotron as a teaching model, but its results showed a skewed stance on China-related topics. The lab ultimately chose Inkling, an open-weight model from Thinking Machines Lab. It hypothesizes that Nemotron's bias came from training alongside Chinese open models, a contamination it says affects most high-quality open models.

The case feeds the debate over political conditioning in Chinese models. A test by The Register on the Qwen3.8-27B-MLX-6bit variant found that, asked about the 1989 events in Tiananmen Square, the model refused or gave evasive answers. The story connects to earlier coverage of Nous Research's first post-training of Qwen3.8-27B, which reported the share of pro-government answers falling from 22% to 0.9%.

On the evaluation side, Hirundo, which builds machine unlearning platforms, published its own experiment on Qwen3.6-35B-A3B: by its figures, output matching Chinese Communist Party positions fell from 89.8% to 2.8% after unlearning.

Nous Research was valued at $1.5 billion in its latest funding round, as reported in October. Arete-27B is available with open weights.