Nous Research says fine-tuning cuts Chinese-state bias from Alibaba's Qwen model

Nous Research, an AI lab best known for its open-source Hermes agent models, said this week that it has fine-tuned Alibaba's open-weight Qwen3.8-27B language model to sharply cut the rate at which it echoes Chinese state talking points on politically sensitive questions. The result, called Arete-27B, is being released free under an Apache 2.0 license, along with the data used to train and test it.

On Nous's own 803-question evaluation set, spanning seven categories such as free expression, democracy and data privacy, the share of answers containing propaganda fell from 22 percent for base Qwen3.8-27B to 0.9 percent for Arete-27B, according to the model card published on Hugging Face. A separate fact score, graded zero to five, rose from 2.32 to 2.98. The swing was sharpest on free-expression questions, where the propaganda rate dropped from 60 percent to 1 percent.

A thin layer, Nous says

Nous's method was deliberately narrow. The team used supervised fine-tuning only, with no reinforcement learning or preference optimization, training on roughly 52.8 million tokens over two epochs and 1,202 steps. A separate model, Thinking Machines' Inkling, served as the teacher generating preferred answers for the fine-tuning set.

The bias is applied late, applied thin, and the fine-tuning removes it without disturbing anything underneath.

General-purpose performance, the company says, barely moved: Arete-27B scored 65.60 percent on Nous's internal Hermes Bench versus 65.77 percent for base Qwen, a gap it calls noise.

Self-graded, and openly so

The evaluation itself came from Nous: both the 803-prompt test set and the grading, done by a separate model acting as judge, are the lab's own design. To let outsiders check its work, Nous published the training data pairs alongside the weights and the held-out evaluation set, an unusually transparent move for a bias claim that is otherwise hard to verify independently.

The release lands amid a wider body of research on political bias in Chinese-developed models. A 2025 NewsGuard analysis found that five leading chatbots, Qwen among them, repeated or failed to correct pro-Beijing disinformation in roughly six of ten tests, and security firm CrowdStrike has reported that DeepSeek produces less secure code when prompts mention Tibet, Taiwan or Falun Gong, according to reporting by The Decoder.

A shift for Nous

The project marks a change in direction for Nous, which built its earlier Hermes models to stay neutral on contested political topics, and explicitly aims instead to reinforce what the company calls Western values. The lab closed a $90 million Series B in October at a roughly $1.5 billion valuation, with backers including Nvidia and Samsung's M12, according to Implicator.ai, which first detailed the release's benchmark caveats. Nous has invited outside researchers to run its evaluation against the released model themselves.

Whether a single round of fine-tuning removes bias as cleanly as Nous describes, or merely suppresses it on the specific prompts in its own test set, is likely to be tested by exactly the kind of independent replication the open release invites.