Models

Nous Research releases Arete-27B, a Qwen3.8-27B fine-tuned on Western values, and says Chinese state bias is a thin layer, a conclusion based on its own 803-prompt test

Nous Research fine-tuned Qwen3.8-27B into Arete-27B and says state bias is superficial; its own 803-prompt test shows propaganda falling from 22% to 0.9%.

By ·

Nous Research announced Arete-27B on November 6, 2026, a version of Alibaba's Qwen3.8-27B post-trained on about 53 million tokens to reinforce what the lab calls Western values. It reports that answers containing propaganda fell from 22% to 0.9% on its own 803-prompt evaluation. The framing needs a limit: this is a fine-tune, not a retraining from scratch, the test and the definition of Western values are Nous's own, and the thin-layer conclusion is the lab's interpretation, not an independent finding.

The essentials

What the test can and cannot show

The 803 prompts, the seven categories and the rubric for what counts as propaganda were all written by Nous, and the main runs use temperature 0 for Arete and the base model but 0.6 for the other models compared. A drop on this test shows the model now answers differently on these prompts; it does not by itself show that the underlying knowledge or behaviour elsewhere is unchanged. AlphaSignal notes that the evaluation measures conformance to a defined value policy rather than a viewpoint-free alignment.

Why it matters beyond Qwen

The Decoder frames the result as evidence that political bias in Chinese open-weight models can be stripped cheaply. The same method can in principle push a model in any direction, which is also the point Nous makes. Nous Research raised 90 million dollars in a Series B in October at a 1.5 billion dollar valuation, per The Decoder.

What is not established

Summary table

Summary table
PointWhat is stated as of November 7, 2026Source
ReleaseAnnounced November 6, 2026; weights, eval and data publicDigg
Base modelAlibaba's Qwen3.8-27B, post-trained on about 53 million tokensThe Decoder
Propaganda rate22% to 0.9% on Nous's 803-prompt testNous Research
General capabilityHermes Bench 65.60% vs 65.77%, an internal testNous Research
LicenseApache 2.0 on Hugging FaceThe Decoder

FAQ

Is Arete-27B a new model?

It is a fine-tune of Alibaba's Qwen3.8-27B, not a model trained from scratch.

Does it prove Chinese AI bias is superficial?

Not independently. It is Nous's own reading of its own test; replication by others has not been seen.

Can I use it?

Nous published the weights, evaluation and training data under Apache 2.0 on Hugging Face, per The Decoder.

Method and limits

Sources checked on November 8, 2026. We read search excerpts of the Nous Research page and X post, not the full blog post or the model card, and relied on The Decoder, Digg, AlphaSignal and Pasquale Pillitteri for details. All figures are Nous's. This is not an endorsement of any political value set.

Sources

Spotted an error in this article? Report it · Our corrections policy

OrigoTech logo

New: the app iPhone & iPad

OrigoTech is now an iPhone and iPad app

Read the articles natively, with “New” markers on unread ones, local search and offline reading of what you have already loaded. Free, no account, iOS 17 or later; ads appear between articles.

Rekha icon: a house drawn in a single stroke

From the same author Self-promotion

An Eulerian path, drawn with your finger

Rekha, the iPhone game by OrigoTech's author: connect every dot in a single stroke without tracing the same link twice, across 500 drawings that are all solvable.

The OrigoTech newsletter

A sourced weekly digest of AI tool news.

Read next