Key points
- Nous Research fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens over two epochs, using supervised fine-tuning only, to produce Arete-27B.
- On Nous's own 803-prompt test, propaganda in answers fell from 22 percent to 0.9 percent; factual accuracy rose from 2.32 to 2.98 out of 5.
- Nous reports its internal Hermes Bench score barely moved, 65.77 percent for the base model versus 65.60 percent for Arete-27B.
- Weights, test dataset and training data are public on Hugging Face under an Apache 2.0 license.
Why it matters for marketers: Teams building on open-weight Chinese models can see that political bias may be correctable with modest fine-tuning, which matters for brand-safe, accurate content on sensitive topics.
Nous Research, the US startup behind the Hermes agent, released Arete-27B, a version of Alibaba's Qwen3.8-27B tuned to remove Chinese state messaging from its answers. Training used supervised fine-tuning on about 53 million tokens, with no reinforcement learning or preference optimization.
Nous measured the result with its own benchmark of 803 prompts across seven categories, including free speech, privacy and surveillance, and democracy. A Grok 4.7 judge checked answers against facts drawn from dated sources. The base model produced propaganda in 22 percent of answers. Arete-27B dropped to 0.9 percent, and the free speech category fell from 60 percent to one percent. These are Nous's own tests, not independent results.
For marketers and content teams, the practical point is reliability. Open-weight models from China rank among the strongest available, and many teams use them for drafting, research and agent workflows. Nous reports refusals, evasive answers and official-line responses on sensitive topics in the base model. The company says general ability held steady, with Hermes Bench scores of 65.60 percent versus 65.77 percent for the base model. Nous argues that the bias is only a thin layer, and says it is applied late in training.
Worth watching is whether outside groups reproduce the numbers, since the weights, test data and training data are public under Apache 2.0. Teams that deploy open models in customer-facing work may want to run their own checks on politically sensitive prompts. Nous also notes that even Nvidia's Nemotron showed some signs of China-related irregularities, so the issue is not limited to Chinese-built models.
Sources
This brief is our own summary. Read the full reporting at the source.