Nous says its Arete-27B fine-tune cut Qwen propaganda answers from 22% to 0.9%

Nous Research has released Arete-27B, a fine-tune of Alibaba’s Qwen3.8-27B that, by Nous’s own measurement, cut answers adopting the Chinese state’s line from 22% to 0.9%. The weights are on Hugging Face under the Apache-2.0 license, the same license as the Qwen base model. The test, the judge and the results all come from Nous, and we found no outside replication yet.

What Nous did

Nous used supervised fine-tuning only: no reinforcement learning, no preference optimization, no architecture changes. It trained all language-model weights on about 52.8 million tokens over two epochs and 1,202 steps, with a teacher model called Inkling supplying the target answers. The post describes the data as 20,000 generated prompts across seven value categories, with overlap against the evaluation set removed.

The model is pitched as a general and agentic assistant for users who want that change in viewpoint. Nous says plainly that it does not claim to have fixed every behavior tied to the original viewpoint.

How the 22% was measured

Nous built an 803-prompt evaluation covering topics such as freedom of expression, state ideology, education and data privacy. Each prompt has two to four required facts from a dated source article. A language-model judge, Grok 4.7, marks each fact as stated, missed or opposed, and flags answers that adopt the official line.

Nous says it tuned the judge against 25 prompts it graded by hand.

That is a judge Nous chose, scoring a rubric Nous wrote, on prompts Nous selected. The headline figure also uses thinking off and temperature 0 for Arete and the base model, while the other models in the table ran at temperature 0.6, which the model card admits are not identical decoding settings.

At temperature 0.6, Arete scores 2.88 on facts against 2.24 for Qwen3.8-27B.

The numbers, as Nous reports them

The fact score, on a 0 to 5 scale, rose from 2.32 to 2.98. Answers scoring 0 fell from 359 to 216 of 803.

The largest shift was in freedom of expression, where the fact score went from 0.89 to 2.44 and the propaganda rate from 60% to 1%. For comparison, the post lists GLM 5.3 at 15% and DeepSeek V4.1 Flash at 16%, and five other models at 0.3% to 0.6%.

A fact score of 2.98 is still below the 3.96 of the teacher model and the 3.65 of Mistral Large 4, which we covered when its open weights shipped. So the fine-tune removes most of the official-line answers, but it does not make the model as factually complete as the best models on this test.

What it cost on general ability

Nous reports 65.60% on Hermes Bench, its internal general-capability benchmark, against 65.77% for the base model. That is a gap of 0.17 points, which Nous calls noise. On Terminal-Bench-2.1 it reports 70% against 67%, but Nous ran both models itself on its own hardware and says it takes no credit for that gain.

Hermes Bench is not public in the sources we opened, so we could not check what it covers. Two benchmarks are a thin basis for the claim of no regression. For how benchmark choices skew rankings, see our explainer on what the big leaderboards measure.

The Decoder, which also covered the release, reports the same figures and the same caveat that Nous ran everything itself.

What to do this week

If your team runs Qwen3.8-27B for anything touching politically sensitive topics, download Arete-27B and run Nous’s published 803-prompt evaluation, then add 50 prompts from your own domain. Judge the answers with a model other than Grok 4.7 as well, and compare against the base model at the same temperature. Treat Nous’s numbers as a hypothesis until your run agrees.

Sources

• Nous Research: Introducing Arete-27B (primary)
• Arete-27B model card on Hugging Face (primary)
• Hugging Face metadata for Arete-27B (license, created 6 November 2026)
• Hugging Face metadata for Qwen3.8-27B (license)
• The Decoder coverage