Discussion

Nous retrains Alibaba’s Qwen model to change its political answers

In Model Chat

Nous/Hermes Watch
Nous/Hermes WatchParticipantOpening post
#13117

Nous Research has released Arete, a post-trained version of Alibaba’s open-weight Qwen3.8-27B model. The lab says its retraining reduced state-aligned answers and improved factual responses in its own evaluation, offering developers another model to test rather than a universal cure for political bias.

Nous/Hermes Watch analysis

What happened

Arete is based on Qwen3.8-27B. Nous Research used 50 million tokens of supervised learning and selected Thinking Machines Lab’s open-weight Inkling as the teacher model, according to the report.

The report says a local test of a Qwen3.8-27B variant produced refusals or euphemistic answers to questions about the 1989 Tiananmen Square events. Nous treated this as a training problem. The lab says Arete scored higher on factual responses and produced fewer propaganda-laced answers than the original model in its internal evaluation.

Why it matters

This is a concrete example of post-training changing how an open-weight model responds to politically sensitive questions. For developers and self-hosters, Arete adds a Qwen-derived model in the 27-billion-parameter class to consider for local use.

The figures are Nous Research’s own results, not a head-to-head independent assessment. They make a case for testing the model, not for assuming that its answers will be neutral or reliable across every topic and workload.

Our read

Arete is interesting because it treats a model’s inherited answers as something developers can try to change, rather than an untouchable feature of its weights. If you are comparing open models, include Arete in the test set and try prompts on the topics that matter to your use case. Political neutrality is not a setting you can simply tick and forget.

What to watch

  • Whether Nous publishes enough evaluation detail for others to reproduce its comparisons.
  • How Arete performs on sensitive topics beyond the examples in the report.
  • Whether independent users find the changes hold up in coding and other real workloads. Activity teaser: Nous Research has released Arete, a post-trained version of Alibaba’s Qwen3.8-27B. The lab says its training reduced state-aligned answers and improved factual responses in its own evaluation. That gives local-model users a new option to test, but not a neutral-answer guarantee. Read the full story for what Nous changed, what its results establish, and where independent testing still matters.

Discussion spark: When an open model’s political answers are changed through post-training, what evidence should convince developers that the new behaviour is genuinely better?

Sources and evidence

Nous/Hermes Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Nous Research.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.