Nous Cut Qwen's Propaganda Rate From 22% to 0.9%. Nous Wrote the Test, Grok Graded It, and the Fact Score Still Trails the Model's Own Teacher.

Nous Research released Arete-27B today, November 6, 2026, at 3:48 PM Central. It is Alibaba's Qwen3.8-27B with a supervised fine-tune and no reinforcement learning: about 52.8 million tokens at 64,000 context, two epochs, 1,202 steps, on roughly 20,000 generated prompts, with Thinking Machines' open Inkling as the teacher. The text data was the only thing trained; the vision encoder stayed frozen and the multi-token-prediction head was disabled. The stated goal is to replace what Nous calls Beijing's framing with "Western values" across seven categories. The license is Apache 2.0.

The scoreboard is Nous's. On an 803-prompt evaluation across the seven categories, graded by Grok 4.7 blind to which model answered, the mean fact score rises from 2.32 to 2.98 out of 5, and the share of answers containing any propaganda falls from 22% to 0.9%. On the freedom-of-expression category the move is larger: fact score 0.89 to 2.44, propaganda 60% to 1%. Three things sit next to those numbers. Nous built the eval and tuned the rubric on a 25-prompt pilot. Arete and Qwen were run at temperature 0 while every other model on the table ran at 0.6, which Nous discloses. And Arete's 2.98 is still below its teacher Inkling at 3.96, Mistral Large 4 at 3.65, Nemotron 3 Ultra at 3.22, and GLM 5.3 at 3.10. We read the model card and the announcement. Here is what held up.

What Nous Built

A viewpoint swap is not neutrality, and Nous does not pretend it is. The card says Western values. Critics called that a different bias rather than the absence of one. Both descriptions are of the same thing; the model answers from a different position, not from no position.

The judge and the exam are the vendor's. Grok 4.7 graded blind, which is good practice. Nous chose Grok, wrote the prompts, and set the rubric, which is also what every lab does with its own benchmark.

The temperature mismatch is in the footnotes. Arete and its base at 0 and the comparison models at 0.6 is a disclosed asymmetry. Nous gives an Arete-at-0.6 row too, 2.88, and Qwen at 0.6 at 2.24, so the gain holds at matched settings even if the headline pair does not.

The Benchmarks, All of Them Nous's

Three honest observations:

• The propaganda number moved the most, and it is the point. From 22% to under 1% on Nous's prompts is what the fine-tune was for, and on that column Arete sits with the Western models.
• The fact score is mid-table. Arete beats its base and Gemma and DeepSeek; it trails four other open models including the one that taught it. A 27B fine-tune did not become a 400B teacher.
• Capability held, which is the quiet result. Hermes Bench barely moved and Terminal-Bench 2.1 ticked up. A political fine-tune that did not break the model is worth noting, with Nous's own disclaimer that it is not taking credit for the coding gain.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Try Arete-27B if you want a Qwen-class open model whose answers on political and historical topics do not carry Beijing's framing, and you are comfortable that the replacement is a stated Western one.

Skip the "most factual" reading; on Nous's own table the most factual open models are its teacher and three others, and the only scoreboard is the vendor's.

The frontier is elsewhere: the model that taught this one, Inkling, scores 3.96 on the same eval and its successor is waiting on its own weights. Our model comparison has the current standings side by side.

Sources

• Hugging Face: NousResearch/Arete-27B
• Nous Research on X: Arete-27B announcement

*Not sure whether an open model's political framing matters for your use? Take the free 2-minute quiz and get matched. Then read about the US open model that led the old board and still has no new score: Muse Spark Open.*