Nous Research Releases Arete-27B, a Qwen3.8-27B Fine-Tune That Reports a 0.9% Propaganda Rate on Its Own 803-Prompt Eval, Graded by Grok 4.7

Overview

Nous Research published Arete-27B on Hugging Face on 2026-11-06 (UTC, per the repository’s commit history). Its model card opens: “We post-trained over Alibaba’s Qwen3.8-27B to reinforce western values on a Chinese-trained model.” Nous reports that the share of answers containing propaganda fell from 22% to 0.9% on an 803-prompt evaluation it built itself, graded by Grok 4.7. Those are Nous’s figures, and none of the sources cited in this article reports an independent reproduction of them.

What Nous released

According to the model card, Arete-27B was trained with supervised fine-tuning only: “No RL, no preference optimization, and no architectural modification.” The card says Nous fine-tuned on about 52.8 million tokens at 64k context over two epochs and 1,202 steps, with Inkling as the teacher model on seven value categories, among them freedom of expression, state ideology and democracy. The model repository’s LICENSE file contains the Apache License, Version 2.0 text. Nous also released an evaluation dataset, Propaganda_Eval, and a training dataset, arete-data-pairs.

The figures Nous reports

The scores below are Nous’s own, from the model card’s table: five runs over all 803 prompts with a 32k-token cap, which the blog post describes as the mean of the five runs. The fact mean runs from 0 to 5. “Any propaganda” is the share of answers scored 1 or 2 on Nous’s framing scale, meaning the answer adopts the Chinese state’s account as its own.

• Arete-27B (thinking off, temperature 0): fact mean 2.98 ± 0.03, any propaganda 0.9%.
• Qwen3.8-27B, the starting model (thinking off, temperature 0): 2.32 ± 0.03, 22%.
• Inkling, the teacher: 3.96 ± 0.07, 0.3%.
• Mistral Large 4: 3.65 ± 0.04, 0.3%.
• GLM 5.3 (thinking off): 3.10 ± 0.06, 15%.
• DeepSeek V4.1 Flash (thinking off): 2.39 ± 0.05, 16%.

The card notes that the other rows used temperature 0.6 and states: “These are not identical decoding settings across all models.” At temperature 0.6 with thinking off, it gives 2.88 ± 0.03 for Arete-27B and 2.24 ± 0.05 for Qwen3.8-27B. In the Herald’s own reading of the card’s table, Arete-27B’s fact mean of 2.98 is below those of five other rows (Inkling, Mistral Large 4, Nemotron 3 Ultra, GLM 5.3 and Muse Glimmer 30B) and above three (Gemma 4 31B, DeepSeek V4.1 Flash and the starting model). Its propaganda rate of 0.9% is above the 0.3% to 0.6% the table gives for five other rows and below the 15% and 16% for GLM 5.3 and DeepSeek V4.1 Flash.

Nous reports that the largest change was in freedom of expression, where the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%. On Terminal-Bench-2.1 it reports 70% for Arete-27B against 67% for the starting model, and says: “We do not take credit for any seeming improvement in the Terminal-Bench-2.1 score.” The blog post adds a thinking-on row for Arete-27B (3.07 ± 0.03, 0.6%) that it marks as provisional, from 2 of 5 runs; the card’s table does not include that row.

What the evaluation measures and its stated limits

The evaluation card states that each prompt carries two to four required facts taken from a source document and that the judge is x-ai/grok-4.7. Nous’s model card says the judge was not told which model wrote an answer, and that the judge’s rubric was revised three times after a 25-prompt pilot that Nous graded by hand. The evaluation card lists its own limitations, including that “Scores depend on the judge model” and that the set is “English only”. It says a prompt was kept only if Grok 4.7 scored 4 or 5 on it, and that Grok 4.7 should not be scored with itself as the judge because it was also that answerability gate.

Where Nous’s own pages differ

The figures in this section are given as each page states them; the Herald has not tried to reconcile them.

• Training data size: the blog post says Nous generated “twenty thousand prompts”, describes the run as “roughly fifty thousand samples”, and captions a chart as covering “28,886 training examples”. The training data card lists 34,878 prompts and 29,236 accepted teacher answers as examples for supervised fine-tuning. The pages cited here do not state how these counts relate.
• Evaluation size: the evaluation card describes 803 prompts, reduced from 834 by removing 31 near-duplicates. The training dataset’s release manifest refers to “the private China censorship eval” and to a final re-check against “all 1,111 eval prompts”. The pages cited here do not say whether that is the same set as the published 803.
• Licences: the evaluation card’s metadata lists license: mit and the model card’s metadata lists license: apache-2.0. The training data card, as read for this article, states no licence. The Decoder, below, writes that the weights, test dataset and training data are all available under Apache 2.0.
• Card edits: by the Herald’s comparison of the repository’s revisions, the model card’s first full version was committed at 19:45 UTC on 2026-11-06 and three further edits followed between 20:51 and 21:04 UTC. They renamed the linked evaluation from Alignment_Eval to Propaganda_Eval and trimmed the release descriptions; the scores are the same in each revision.

Independent reporting and context

The Decoder reported the release on 2026-11-07, crediting Nous Research as its source. Its account of the results does not describe a reproduction.

In a separate report dated 2026-11-01, The Decoder described a study by Vals AI of how Chinese and other models answer politically sensitive questions about China. It reports wide variation between models, and that for most models answers in Simplified Chinese were more pro-government than in English, Russian or Spanish. It also reports that Vals scored responses with GPT-5.6 Luna, that refusals counted as neutral by default, so that a model blocking every question could appear neutral, and that many Chinese models were tested through third-party providers whose results may differ from the manufacturers’ own services. Vals used a different judge and scale from Nous, and the Herald does not compare the two sets of numbers. In Nous’s evaluation, according to its model card, unanswered prompts count as 0 on the fact score.

What is not yet known

• None of the cited sources reports a reproduction of the Arete-27B scores by a party other than Nous, or scores from a judge other than Grok 4.7.
• The evaluation card says the set is English only. The cited sources do not report how Arete-27B behaves in other languages, although The Decoder’s account of the Vals study says the prompt language affected answers.
• None of the cited sources includes a response from Alibaba or the Qwen team.
• The model card says the model is “intended for general assistance and agentic work where users want this change in viewpoint”, and says of the results: “That is not to claim that we remedied all behaviors associated with the extant viewpoint, but that we have addressed a potential root cause.”