Nous Retrained Qwen to Cut ‘Propaganda’ From 22% to 0.9%, but Its Judge Is Grok 4.7
Nous Research fine-tuned Qwen3.8-27B into Arete-27B, reporting a drop in propaganda answers from 22% to 0.9% on its own 803-prompt test. The weights are Apache 2.0, but the judge was Grok 4.7 and no outside lab has replicated the result.
Nous Research published Arete-27B on 6 November, a supervised fine-tune of Alibaba’s Qwen3.8-27B that it says cut answers containing Chinese state propaganda from 22% to 0.9%. The model keeps the base weights’ 27B size and ships under Apache 2.0, so anyone can download, run or fine-tune it again.
The recipe is deliberately small: 20,000 generated prompts, 28,886 training examples, about 52.8 million tokens and 1,202 steps. Nous says it used supervised fine-tuning only, with no reinforcement learning, preference optimisation or architecture change.
The catch sits in the measurement, not the training. The 803-prompt evaluation behind the headline number is Nous’s own, the judge is Grok 4.7, and no outside lab has replicated the result. The eval measures one kind of bias, the kind that echoes Beijing, and says nothing about bias in the other direction.
Quick answer: Arete-27B is a 27B open-weight model from Nous Research, built by fine-tuning Alibaba’s Qwen3.8-27B on about 52.8 million tokens of teacher-generated answers. Nous reports that its fact score rose from 2.32 to 2.98 out of 5 and that answers containing Chinese state propaganda fell from 22% to 0.9% across 803 prompts. Both numbers come from Nous’s own evaluation, judged by Grok 4.7. The weights are Apache 2.0. The result shows that value behaviour can be changed cheaply, not that the new values are correct.
What Nous actually changed
Nous picked Qwen3.8-27B because it was the strongest model that would run locally and because it showed the behaviour the lab wanted to target. It says it saw refusals, deflections and answers framed in the language of the Chinese Communist Party on politically sensitive topics.
It built an 803-prompt evaluation across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. Each prompt carries two to four required facts from a dated source, plus the official CCP framing on that topic.
The judge, Grok 4.7, gets the prompt, the answer, the required facts and a fixed rubric, but not the model name. A fact is marked stated, missed or opposed, and an answer that adopts the official line as its own is marked as propaganda. Nous built the rubric on a 25-prompt hand-graded pilot and revised it three times where judge and human disagreed.
The teacher was Inkling, the open model from Thinking Machines, picked after Nous considered Nemotron 3 Ultra and found what it called a checkered record on China-related topics. Nous says it used an open teacher partly to avoid restrictions on distilling a closed model.
The final set was 28,886 examples at 64k context, more than the 20,000 prompts because the pipeline kept multiple responses per prompt. Nous published the pairs as arete-data-pairs and the held-out evaluation separately.
The numbers, and the asterisks
Nous reports a fact score of 2.98 out of 5 for Arete-27B against 2.32 for Qwen, with propaganda answers falling from 22% to 0.9%. The blog rounds that to 1%, while the model card keeps 0.9%. Answers scoring zero, meaning no required fact stated or at least one fact opposed, fell from 359 to 216 out of 803.
| Model | Fact score (0 to 5) | Answers with propaganda | Note |
|---|---|---|---|
| Inkling (teacher) | 3.96 | 0.3% | Best fact score in the table |
| Mistral Large 4 | 3.65 | 0.3% | Closed benchmark entry |
| Nemotron 3 Ultra | 3.22 | 0.6% | Thinking off |
| GLM 5.3 | 3.10 | 15% | Higher facts, far more propaganda |
| Arete-27B (thinking on) | 3.07 | 0.6% | Provisional, 2 of 5 runs |
| Muse Glimmer 30B | 3.02 | 0.5% | Thinking off |
| Arete-27B (thinking off) | 2.98 | 0.9% | Temperature 0 |
| Gemma 4 31B | 2.58 | 0.5% | Thinking off |
| DeepSeek V4.1 Flash | 2.39 | 16% | Thinking off |
| Qwen3.8-27B (base) | 2.32 | 22% | Temperature 0 |
Source: Nous Research, 6 November 2026. Each model answered all 803 prompts five times with a 32k-token cap. Standard deviations were between 0.02 and 0.07.
Two details matter more than the headline. First, Arete and its Qwen base ran at temperature 0 while most other models ran at 0.6, so the cross-model comparison is not perfectly like for like. At 0.6, Nous says Arete scored 2.88 and Qwen 2.24, so the improvement survives.
Second, the fact score is not a general capability score. GLM 5.3 states more required facts than Arete, 3.10 against 2.98, yet 15% of its answers still adopt the official line. Arete wins on framing, not raw recall, and still trails the teacher, Mistral Large 4 and Nemotron 3 Ultra on facts.
Capability held, with one caveat
Nous ran two capability checks. On Hermes Bench, its internal general-capability test, Arete scored 65.60% against 65.77% for Qwen, a gap the company calls noise. On Terminal-Bench-2.1 it scored 70% against 67%, and Nous declines to take credit because inference configurations and harnesses differed.
The caveat is context. Qwen3.8-27B supports 262,144 tokens natively, but Nous trained Arete at 64k with no sequence packing. The model card does not claim the full window survived, and long-context behaviour was not evaluated.
The thinking-on result is also provisional: 3.07 and 0.6% propaganda from two of five runs, to be updated when the rest finish. The stable row is thinking off at 2.98 and 0.9%.
Confirmed vs not confirmed
| Claim | Status | Evidence |
|---|---|---|
| Arete-27B is a fine-tune of Qwen3.8-27B | Confirmed | Model card and Hugging Face metadata name the base model |
| Weights are Apache 2.0 | Confirmed | License tag on the model repository |
| Training used about 52.8M tokens and 1,202 steps | Confirmed by Nous | Blog and model card give the same figures |
| Propaganda rate fell from 22% to 0.9% | Confirmed as a Nous result | Nous evaluation, 803 prompts, Grok 4.7 judge |
| The drop generalises to other topics or languages | Not confirmed | Eval covers seven English-language categories only |
| An independent lab reproduced the result | Not confirmed | No replication published as of 8 November |
| The bias layer is shallow in every open model | Not confirmed | Nous tested one base model and one teacher |
| Removing the refusals made the model safer overall | Not confirmed | No red-team or harmful-content evaluation published |
Why “Western values” is a product decision
The framing is a choice, not neutral. Nous chose the categories, wrote the rubric and picked the judge, so a different lab could build a different prompt set and get a different ranking. The prompts, judge prompt and scoring scripts are public, so the test can be argued with rather than taken on trust.
The release does not answer the safety question. A model that stops deflecting on political topics may also stop refusing where a refusal was appropriate. Nous reports no harmful-content or jailbreak evaluation, and its capability checks cover reasoning and agentic work, not misuse resistance.
Alibaba’s Apache 2.0 licence permits the fine-tune, so Nous needed no permission. That is the quiet point: once a strong model ships open weights, its value layer is a starting point, not a fixed feature. The same logic applies to every Chinese open-weight release, and to every American one.
What it means for Indian developers
A 27B dense model under Apache 2.0 is practical for Indian teams that want to fine-tune for Indic languages or local policy positions without asking a US lab for access. Nous did not disclose GPU hours or dollar cost, so the budget case is still unproven.
The harder lesson is evaluation. An Indian team that fine-tunes Arete for Hindi or Tamil cannot inherit the 0.9% number, because the published test is English. It needs its own held-out set, human spot checks and a judge outside its own model family.
Distribution is already moving. Community quantisations appeared on Hugging Face within a day, including a GGUF build with more than 5,000 downloads at the time of writing. A 27B model needs roughly 54 GB at 16-bit before overhead, so most local deployments will use 8-bit or 4-bit weights and accept some quality loss. That is a workstation or single server GPU, not a laptop.
The earlier Qwen weight-editing result showed that a narrow intervention can move one benchmark without changing the model underneath. Arete is a cleaner test because the intervention is ordinary supervised fine-tuning and the evaluation is published with it.
Frequently asked questions
What is Arete-27B?
An open-weight model from Nous Research, made by post-training Alibaba’s Qwen3.8-27B. It is designed to answer questions about Chinese politics and society with facts rather than the state’s framing, and it ships under Apache 2.0.
How much did it cost to train?
Nous has not published a dollar figure or GPU-hour count. The disclosed inputs are about 52.8 million tokens, 28,886 examples, two epochs and 1,202 steps. Any cost estimate beyond that is a guess.
Is Arete-27B free to use?
The weights are free and the licence permits commercial use and modification. Running it is not free: you still pay for GPUs or a hosted endpoint, and most teams will rent capacity.
Does it remove all bias from Qwen?
No. It changes behaviour on seven categories measured by 803 English prompts. It does not test every topic, language or kind of bias, and the evaluation was designed by the lab that trained the model.
Sources
- Nous Research, “Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values,” 6 November 2026.
- Hugging Face, Arete-27B model card and metadata, accessed 8 November 2026.
- Hugging Face, Propaganda_Eval dataset, accessed 8 November 2026.
- Hugging Face, arete-data-pairs training dataset, accessed 8 November 2026.
- Hugging Face, Qwen3.8-27B model card, accessed 8 November 2026.
- Hugging Face API, Qwen3.8-27B download count, accessed 8 November 2026.