---
title: "Arete-27B - turning Qwen3.8 from a 22% propaganda rate to 0.9%, with open weights and data"
description: "Nous Research post-trains Alibaba's Qwen3.8-27B to reinforce Western values. An 803-prompt eval shows propaganda drops from 22% to 0.9%. Hermes Bench shows no significant capability tradeoff."
date: 2026-11-07
reading_time: "4 min read"
url: https://myargoquest.com/blog/arete-27b
markdown: https://myargoquest.com/blog/arete-27b.md
publisher: Argo (myargoquest.com)
---

# Arete-27B - turning Qwen3.8 from a 22% propaganda rate to 0.9%, with open weights and data

# Arete-27B: turning Qwen3.8 from a 22% propaganda rate to 0.9%

Nous Research released Arete-27B, a post-trained version of Alibaba's Qwen3.8-27B that reduces Chinese state-aligned behavior from 22% to 0.9% while retaining capability. The weights, evaluation and training data are all open source, and the methodology is documented end to end.

This is not an uncensoring project. It is a values alignment project, and the distinction matters.

![A horizontal bar chart comparing six models: Qwen base 22%, DeepSeek 16%, GLM 15%, Arete 0.9%, Mistral and Inkling below 1%](https://myargoquest.com/blog/arete-27b-1.jpg)
*Arete-27B drops propaganda from 22% to 0.9%, in line with Western models.*

## What the benchmark measures

Nous built an evaluation of 803 prompts across seven value categories: freedom of expression, democracy, state ideology, state versus free market, education, data privacy and surveillance, and the environment. For each prompt they extracted two to four required facts from a dated primary source and the official Chinese Communist Party framing language for that topic.

A Grok 4.7 judge, not told which model wrote the answer, received the prompt, the model's answer, the required facts and a fixed rubric. It recorded for each fact whether the answer stated, missed or contradicted it, and quoted any phrase where the answer adopted the official line. The judge was developed on a 25-prompt pilot graded by hand, with the rubric revised three times wherever the judge and the human grades disagreed.

The key numbers from the eval table:

| Model | Fact mean (0-5) | Propaganda rate | |---|---|---| | Inkling (teacher model) | 3.96 | 0.3% | | Mistral Large 4 | 3.65 | 0.3% | | Arete-27B (thinking off) | 2.98 | 0.9% | | GLM 5.3 | 3.10 | 15% | | DeepSeek V4.1 Flash | 2.39 | 16% | | Qwen3.8-27B (base) | 2.32 | 22% |

The propaganda rate drops by a factor of 24. The fact accuracy score improves from 2.32 to 2.98. The Hermes Bench shows no significant regression.

![A bar chart showing fact accuracy scores: Inkling 3.96, Mistral 3.65, Arete 2.98, Qwen base 2.32](https://myargoquest.com/blog/arete-27b-2.jpg)
*Fact accuracy improves alongside the propaganda drop. Arete scores 2.98 against Qwen base 2.32.*

## How it was built

The training uses about 53 million tokens of synthetic data, generated by a teacher model called Inkling. The exact training recipe is not the story here; what matters is that the eval, the data and the weights are all published, so the result is reproducible and improvable by anyone.

The blog makes an explicit point about why this matters beyond the specific model: open-weight Chinese models are among the strongest available, but they ship with state-aligned behavior in the weights. The US government has taken steps to restrict Chinese AI from federal use, and studies have shown that geopolitical information in context can affect code quality.

![A five-step pipeline from Qwen3.8 base to Arete-27B: eval, dataset, training, result](https://myargoquest.com/blog/arete-27b-3.jpg)
*How Arete-27B was built: base model, eval, training data, fine-tuning, result.*

## The seven value categories

The evaluation covers seven categories, each designed to test a dimension where Chinese state-aligned framing conflicts with Western values: freedom of expression, democracy, state ideology, state versus free market, education, data privacy and surveillance, and the environment.

The categories were chosen to cover the main axes where political bias in an open model would have tangible downstream consequences, not just in political answers but in everyday reasoning where a model's training viewpoint leaks into factual decisions.

![Seven boxes naming the eval categories: freedom of expression, democracy, state ideology, state vs free market, education, data privacy, and environment](https://myargoquest.com/blog/arete-27b-4.jpg)
*The evaluation covers seven value categories where Chinese state-aligned framing conflicts with Western values.*

## What it means

Arete-27B is a demonstration that the values in an open model are one fine-tuning away from changing. The blog says it plainly: "Open-weight Chinese models are among the strongest available, but they ship with state-aligned behavior baked into the weights. The values in an open model are one fine-tune away."

The implication is not that all Chinese models should be replaced. It is that openness gives labs the raw tools to mold intelligence to varied preferences, and that the current model landscape has a homogeneity of viewpoint that runs through frontier labs regardless of geography.

## What is not claimed

The blog is careful about limits. It says "we have addressed a potential root cause," not that every CCP-aligned behavior is gone. The eval judge was developed on a 25-prompt pilot with three rubric revisions, which is thorough for a research project but not an industrial certification. The model's Hermes Bench score shows no significant tradeoff, but individual task categories may differ.

## Open source

All assets are public: - Weights: Arete-27B on HuggingFace - Eval: Propaganda_Eval (803 prompts, 7 categories) - Training data: arete-data-pairs (about 53 million tokens)

The full blog post with detailed methodology: nousresearch.com/arete

![Three panels showing the open source components: weights, evaluation, training data](https://myargoquest.com/blog/arete-27b-5.jpg)
*Weights, eval and training data: all open source, all reproducible.*
