Nous Research released Arete-27B, an open-weight supervised fine-tune of Alibaba's Qwen3.8-27B model designed to strip Chinese state propaganda framing. Trained on 53 million tokens using Thinking Machines' Inkling model as a teacher, Arete-27B reduced propaganda responses on an 803-prompt evaluation from 22 percent to 0.9 percent without sacrificing general benchmark capability. Independent research from NewsGuard, CrowdStrike, and Aleph Alpha highlights widespread state-aligned bias across major Chinese AI models.
Arete-27B release and fine-tuning methodology
- ▪Nous Research selected Inkling, an open model from Thinking Machines Lab, as the teacher model for Arete-27B to avoid terms-of-use restrictions on distillation
- ▪Nous Research released Arete-27B, an open-weight supervised fine-tune of Alibaba's Qwen3.8-27B designed to remove Chinese state framing from model responses
- ▪The training dataset for Arete-27B contains 34,878 prompts across seven value categories, with 29,236 accepted teacher answers forming the fine-tuning split
- ▪Nous Research trained Arete-27B using supervised fine-tuning on approximately 53 million tokens over two epochs without reinforcement learning or preference optimization
Evaluation results on propaganda reduction and accuracy
- ▪Nous Research evaluated Arete-27B using xAI's Grok 4.7 as a judge model to grade 803 prompts across seven value categories against facts from cited sources
- ▪On an 803-prompt evaluation set created by Nous Research, Arete-27B reduced the proportion of responses containing Chinese state propaganda from 22 percent to 0.9 percent
- ▪On the 803-prompt Propaganda_Eval benchmark, answers scoring zero on factual accuracy dropped from 359 for Qwen3.8-27B to 216 for Arete-27B
- ▪Arete-27B raised the factual accuracy score on the Nous Research benchmark to 2.98 out of 5, compared to 2.32 out of 5 for base Qwen3.8-27B
- ▪In the freedom of expression evaluation category, Arete-27B reduced the propaganda response rate from 60 percent to 1 percent while raising the factual accuracy score from 0.89 to 2.44
Findings on Chinese model bias
- ▪Based on the fine-tuning results of Arete-27B, Nous Research stated that the state alignment bias layer in Chinese language models is shallow and applied late in training
- ▪Nous Research reported observing refusals, evasive answers, and responses aligned with the Chinese Communist Party on sensitive topics in base Qwen3.8-27B
Broader Research on bias in Chinese AI models
- ▪A study by CrowdStrike found that DeepSeek generates less secure code when user queries reference topics such as Falun Gong, Tibet, or Taiwan
- ▪An Aleph Alpha study analyzing 967 politically sensitive questions found that between 17 percent and 41 percent of responses from Alibaba, DeepSeek, and Moonshot AI models were balanced
- ▪A 2025 NewsGuard study found that five leading Chinese AI models, including Qwen, repeated or failed to correct pro-Chinese disinformation in 60 percent of test cases
Nous Research funding and expansion
- ▪Nous Research plans to use its Series B funding to expand into enterprise sales with Hermes for Businesses
- ▪In October 2026, Nous Research raised $90 million in Series B funding led by Robot Ventures at a $1.5 billion valuation, with participation from Nvidia and Samsung
Story comments
Loading comments…