Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9%

What happened

Nous Research released Arete-27B, a modified version of Alibaba's Qwen3.8-27B language model, after supervised fine-tuning to reduce Chinese state propaganda in its responses. The new model was evaluated using an 803-prompt test, where the rate of propaganda in answers dropped from 22% to 0.9%. Arete-27B maintains similar performance to the original model on capability benchmarks. The model weights, evaluation set, and training data are all publicly available under open licenses.

Key details

• Arete-27B has 27 billion dense parameters and uses a 64k context window for training.
• The propaganda rate on the evaluation set fell from 22% in Qwen3.8-27B to 0.9% in Arete-27B.
• Fact recall score increased from 2.32 to 2.98 out of 5 on the same evaluation.
• The training process used 52.8 million tokens, two epochs, and a batch size of 48.
• Arete-27B is released under the Apache 2.0 license, and the evaluation set is MIT licensed.

Why it matters

Developers who want to use a large language model with Qwen3.8-27B’s capabilities but without Chinese state-aligned responses now have an alternative. The open release of both the model and its training data allows researchers and companies to inspect, adapt, or deploy the model without restrictions. This change directly affects those concerned about content alignment and transparency in AI systems, giving them a model with substantially reduced state propaganda in its outputs.