THE CRUNCH

US startup Nous Research says a modest fine-tuning run can largely strip Chinese state propaganda from an open-weight model. Starting from Alibaba's Qwen3.8-27B, the team trained a new open-weight model called Arete-27B on roughly 53 million tokens using supervised fine-tuning, with no reinforcement learning or preference optimisation. According to Nous, the share of answers containing propaganda fell from 22 percent to 0.9 percent, while factual accuracy on its own test rose from 2.32 to 2.98 out of 5.

The company built its own benchmark of 803 prompts across seven categories, including free speech, state ideology, privacy and surveillance, and democracy, each tied to facts from dated source articles. Grok 4.7 from xAI served as the judge model, checking blind whether a response mentioned, omitted or contradicted those facts. The biggest change came in the free speech category, where the propaganda rate dropped from 60 percent to one percent.

Nous says general capabilities held up: on its internal Hermes Bench, Arete-27B scored 65.60 percent against 65.77 percent for the base model, and Terminal-Bench 2.1 rose from 67 to 70 percent, though the team does not count that as a real gain because the test setups differed. The team concludes that the bias layer in a Chinese-aligned model is shallow, applied late and thin, and removable with carefully chosen supervision without disturbing anything underneath.

The finding fits earlier research: in 2025, NewsGuard found five leading models including Qwen repeated or failed to correct pro-Chinese disinformation in 60 percent of cases, and a study by Aleph Alpha using 967 politically sensitive questions found only 17 to 41 percent of answers from Alibaba, DeepSeek and Moonshot AI models were balanced. A ValsAI study also showed Chinese models are not uniform, with GLM-5.3 Flash confirming human rights violations while Kimi K3 denied the existence of internment camps.

One wrinkle in the training pipeline: Nous chose Inkling from Thinking Machines Lab as its teacher model because open models would not break terms of service banning distillation, but Inkling's own synthetic training data was partly generated with the Chinese Kimi K2.5, and Nous reports similar contamination in its own projects. Nvidia's Nemotron 3 Ultra was also shortlisted but rejected after what Nous describes as signs of irregularities on China-related topics.