Nous Research fine-tunes Qwen3.8-27B to cut Chinese state propaganda

Nous Research fine-tuned Alibaba's Qwen3.8-27B into Arete-27B and cut the model's propaganda rate from 22 percent to under one percent, while keeping general capabilities intact. The company says the political bias in the Chinese open-weight model was only a thin layer that supervised fine-tuning on 53 million tokens could remove. Arete-27B, its weights, test dataset and training data are available under an Apache 2.0 license.

Testing propaganda and factual accuracy

Nous built an 803-prompt test across seven categories, including free speech, state ideology, education, privacy and surveillance, democracy and the environment. For each prompt, the team defined two to four facts from dated source articles. Grok 4.7 served as judge model, checking whether responses mention, omit or contradict those facts. Unmodified Qwen3.8-27B scored 2.32 out of 5 on factual accuracy and 22 percent of answers contained propaganda. Western models such as Mistral Large 4 and Gemma 4 31B scored below one percent, while GLM-5.3 and DeepSeek V4.1 Flash scored 15 and 16 percent.

Training and results

Nous used supervised fine-tuning on about 53 million tokens over two epochs, without reinforcement learning or preference optimization. It chose Inkling from Thinking Machines Lab as teacher model because an open model would not break terms of service banning distillation; Nvidia's Nemotron 3 Ultra was also considered but showed signs of irregularities on China-related topics. Arete-27B scored 2.98 on factual accuracy and 0.9 percent propaganda. In free speech, the fact score rose from 0.89 to 2.44 and propaganda fell from 60 to one percent. With reasoning mode on, Arete reached 3.07, though based on only two of five runs. General capabilities stayed near base: Hermes Bench 65.60 versus 65.77 percent; Terminal-Bench 2.1 rose from 67 to 70 percent, though Nous does not call that a real gain because test setups differed.

Context and company

Nous says the bias layer in a Chinese-aligned model is shallow, applied late and thin, and that about 50 million tokens of carefully chosen supervision removes it without disturbing what is underneath. The startup raised $90 million in a Series B round in October at a $1.5 billion valuation, led by Robot Ventures with Nvidia and Samsung among participants. It is moving into enterprise sales with 'Hermes for Businesses.' Other studies document the same bias: NewsGuard found five leading models repeated or failed to correct pro-Chinese disinformation in 60 percent of cases in 2025; CrowdStrike finding DeepSeek produces less secure code when queries reference Falun Gong, Tibet or Taiwan; Aleph Alpha's study of 967 politically sensitive questions found 17 to 41 percent balanced answers among Alibaba, DeepSeek and Moonshot AI models; and a ValsAI study found GLM-5.3 Flash confirmed human rights violations while Kimi K3 denied internment camps.

Sources: The Decoder