Nous Research strips Qwen3.8-27B propaganda to under 1% with Arete-27B

The release comes as Chinese open-weight models rank among the strongest available, but according to Nous they carry state-mandated behavior baked into their weights. With Qwen3.8-27B, the team observed refusals, evasive answers, and responses aligned with the Chinese Communist Party. To measure this, Nous built a test with 803 prompts across seven categories, using Grok 4.7 as the judge model to check whether responses mention, omit, or contradict predefined facts.

Nous chose Inkling from Thinking Machines Lab as its teacher model because an open model wouldn't break terms of service that ban distillation. The team also considered Nvidia's Nemotron 3 Ultra but noted signs of irregularities on China-related topics, suspecting the bias carried over from training on Chinese open-weight models. The training data for Arete-27B is spread nearly evenly across the seven value categories, with most examples running under 1,000 tokens.

The result adds to a growing body of evidence on Chinese model bias. In 2025, NewsGuard found that five leading models, including Qwen, repeated or failed to correct pro-Chinese disinformation in 60 percent of cases. CrowdStrike found that DeepSeek produces less secure code when queries reference Falun Gong, Tibet, or Taiwan. A study by Aleph Alpha with 967 politically sensitive questions reached similar conclusions: among models from Alibaba, DeepSeek, and Moonshot AI, only 17 to 41 percent of answers were balanced. Nous Research itself raised $90 million in a Series B round in October at a $1.5 billion valuation, with Robot Ventures leading and Nvidia and Samsung among the participants.