Nous Research releases Arete-27B, a Qwen3.8-27B fine-tune that cuts measured Chinese state propaganda in answers from 22% to 0.9%

Nous Research released Arete-27B, an open-weight model under Apache 2.0 on Hugging Face. It is Alibaba's Qwen3.8-27B fine-tuned on about 53M tokens over two epochs, using supervised fine-tuning only, with no reinforcement learning or preference optimisation. Nous also published the training data and a new 803-prompt, seven-category test set. On Nous's own benchmark, judged by Grok 4.7, the rate of Chinese state propaganda in answers fell from 22% to 0.9%, factual accuracy rose from 2.32 to 2.98 out of 5, and its general Hermes Bench score barely moved (65.60% vs 65.77%). We are reading The Decoder's write-up, not Nous's post.

Anyone deploying Qwen-family models, such as enterprises, developers and researchers, now has a documented and cheap recipe, plus public data, for reducing one kind of political skew. Because the data and test set are open, others can rerun and challenge the result, which is the most useful part of this release. The impact is moderate. It is a fine-tune of someone else's base model from a smaller lab, and it only shows that this approach works on this benchmark, not that it works on other models or in real use.

The "thin layer" framing is a broad conclusion from one self-built benchmark, scored by a single judge model (Grok 4.7), so the 22% to 0.9% figure is unverified and could partly reflect the test being matched to the fix. "Propaganda" is also a contested label, and removing it may shift answers toward the views of whoever curated the training data and wrote the judge's rubric. No independent evaluation exists yet, and the near-flat Hermes Bench score is also Nous's own measure.

1. 01Independent replications on the released 803-prompt set and on other test sets, using different judge models than Grok 4.7, to see if the 0.9% rate holds.
2. 02Review of the published training data for bias in the opposite direction, and checks on whether Arete-27B has lost refusals or accuracy on non-political tasks.
3. 03Whether others apply the same method to different base models (for example DeepSeek V4.1 Flash or GLM-5.3), and whether Alibaba or Hugging Face leaderboards and download figures show real adoption over the next few weeks.