Nous Research's Arete-27B Strips Chinese State Propaganda From Qwen Down to 0.9%

Nous Research has released Arete-27B, a supervised fine-tune of Alibaba’s Qwen3.8-27B designed to change how the model handles politically sensitive topics. The weights are available on Hugging Face under Apache 2.0.

On Nous’s 803-prompt benchmark, Arete’s mean factual score rose from 2.32 to 2.98 out of 5, while answers classified as containing propaganda fell from 22% to 0.9%. Two general capability benchmarks showed no meaningful regression, suggesting that targeted supervised fine-tuning can alter a specific behavior without broadly degrading the underlying model. All results remain self-reported and await independent replication.

Strong weights, unwanted framing

Chinese open-weight models rank near the top of several open-model leaderboards and increasingly power local deployments. Nous reports that Qwen3.8-27B sometimes refused, deflected, or adopted Chinese state framing when answering questions about politics, policy, and history.

The project tests whether that behavior can be changed through a relatively small amount of targeted post-training. Nous’s hypothesis is that the relevant framing can be adjusted through supervised examples while preserving the capabilities learned during pretraining.

An eval that separates facts from framing

Before training Arete, the team created an evaluation with 803 prompts across seven categories:

• Freedom of expression
• State ideology
• State control versus free markets
• Education
• Data privacy and surveillance
• Democracy
• Environmental policy

The benchmark scores two distinct properties that broad bias evaluations often combine:

• Factual recall: Whether an answer includes two to four required facts extracted from a dated source article, producing a score from 0 to 5.
• Political framing: Whether an answer presents the Chinese state’s account as its own by replacing a required fact with the official position or endorsing state conduct while recounting the facts.