Nous Research stripped Chinese state propaganda out of Qwen with a thin layer of fine-tuning

Chinese open-weight models are among the best you can download, with Beijing's talking points included. Nous Research says removing them takes surprisingly little. Its new model, Arete-27B, is Alibaba's Qwen3.8-27B fine-tuned to drop state propaganda, Jonathan Kemper reports for The Decoder.

How they measured it

Nous wrote 803 test prompts across seven areas, from surveillance and privacy to free expression, democracy, education, official ideology and the environment. Each prompt came with two to four facts from a dated source. A judge model, xAI's Grok 4.7, checked whether answers mentioned, omitted or contradicted those facts, and counted an answer as propaganda if it repeated, endorsed or treated Beijing's narrative as equally valid. The judge didn't know which model wrote each answer.

The base Qwen3.8-27B scored 2.32 of 5 on factual accuracy, with propaganda in 22% of answers. Western models like Mistral Large 4 and Gemma 4 31B came in under 1%; two other Chinese models, GLM-5.3 and DeepSeek's V4.1 Flash, landed at 15% and 16%.

The fix

The training itself was plain supervised fine-tuning, two passes over roughly 53 million tokens, with no reinforcement learning. Its teacher model was Thinking Machines' Inkling, chosen partly because using an open model avoids terms that ban distillation. Nous also looked at Nvidia's Nemotron 3 Ultra, but it showed irregularities on China topics, which Nous suspects came from training on Chinese open models.

Arete-27B scores 2.98 on accuracy, with propaganda down to 0.9%. In free speech, the propaganda rate fell from 60% to 1%. General ability held: 65.60% on Nous's Hermes Bench versus 65.77% for the base model. Nous's conclusion is that the bias is "applied late" and "applied thin."

Caveats

Nous wrote the benchmark, picked the judge and reports its own results, and "Western values" is an explicit goal, not neutrality, which is a shift from its 2024 Hermes 3. Weights, test set and training data are on Hugging Face under Apache 2.0, so others can check.

For developers, the practical point is that a model's values are shallower than its capabilities: a modest, targeted fine-tune changes what it will say without changing what it can do. That cuts both ways.

Scoreboard update: Draw. A small fine-tune scrubs one government's talking points out of a model, and could just as easily add another's.

Stay undead.