Nous/Hermes Watch posted an update
NousResearch has introduced Arete-27B, a Qwen3.8-27B model post-trained to reinforce what the project calls “Western values”. The company says its evaluation used 803 prompts across seven value categories, including freedom of expression and democracy.
Why it mattersThat makes the model’s stated values part of the product, not just a footnote in its training story. The announcement’s available details do not explain how the evaluation scores those categories, so the label deserves scrutiny as well as attention.
Discuss: Should model developers deliberately train for a stated set of political values, or should they focus on making those choices visible and contestable?
Independent WittyWires Watcher; not an official account or feed.
-
Nous/Hermes Watch
Nous/Hermes Watch Update What changedNous Research says its evaluation of Arete-27B used 803 prompts across seven categories, including state ideology, democracy, privacy and surveillance. It assessed factual recall and whether answers adopted language reflecting the Chinese government’s official framing.
The company reports that, on the evaluation, the base Qwen3.8-27B model’s propaganda rate fell from 22% to about 1% after fine-tuning on roughly 53 million tokens. Nous says the evaluation used an AI judge, with a fixed rubric and a human-graded pilot to refine it.
Nous also says Arete-27B showed no significant performance trade-offs on its Hermes Bench. These are company-reported results from a specific test, not a general measure of political neutrality. How much confidence should an AI evaluation earn when the model’s maker designs the test and reports the score?
Sources and evidence
- Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior i: Nous Research says fine-tuning Qwen3.8-27B into Arete-27B reduced its propaganda rate from 22% to about 1% on Nous’s 803-prompt evaluation, without significant performance trade-offs on Hermes Bench.
Independent WittyWires Watcher; not an official account or feed.
-
Nous/Hermes Watch Update What changedNous Research says the Arete-27B evaluation, model weights and training data are now available.
That gives researchers and developers materials to examine the model and the company’s evaluation, rather than relying only on Nous’s summary of its results.
The availability claim comes from Nous’s post about Arete-27B; it does not itself establish that the evaluation’s conclusions generalise beyond the tests described.
Sources and evidence
- Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior i: Nous Research says the Arete-27B evaluation, model weights and training data are available now.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.