Nous Research's Arete-27B cuts Qwen's propaganda rate from 22% to 0.9%

Nous Research published Arete-27B on 7 November 2026, an open-weight model it made by fine-tuning Alibaba's Qwen3.8-27B to remove what it calls CCP-framed answers on politically sensitive topics. The Hugging Face model card, under an Apache 2.0 licence, says the answer rate for propaganda fell from 22% to 0.9% and the fact score rose from 2.32 to 2.98 out of 5. The Decoder reported the same figures the same day.

How Nous measured it

Nous built its own test: 803 prompts across seven categories, including free expression, state ideology, education, data privacy and surveillance, and democracy. Each prompt carries two to four required facts taken from a dated source article. An LLM judge, xAI's Grok 4.7, marks each fact stated, missed or opposed, and scores whether the answer adopts the Chinese state's account as its own. The judge is not told which model wrote the answer.

The model card says the rubric was revised three times against a hand-graded 25-prompt pilot. A single opposed fact scores 0 on the fact measure, whatever else the answer says.

What changed, and what did not

Per The Decoder, the biggest shift was in free expression: the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%. Nous reports Western models such as Mistral Large 4 and Gemma 4 31B score under 1%, against 15% for GLM-5.3 and 16% for DeepSeek V4.1 Flash.

The training recipe is plain. Nous used supervised fine-tuning only, with no reinforcement learning or preference optimisation, on about 52.8 million tokens at 64k context over 1,202 steps. It generated 20,000 prompts, removed any that overlapped the test, and used Thinking Machines' open Inkling model as the teacher to avoid terms-of-use limits on distillation. On general ability, Nous reports 65.60% on its own Hermes Bench against 65.77% for the base model. Terminal-Bench 2.1 moved from 67% to 70%, which Nous does not call a real gain because the test setups differed.

What it leaves out

Both the test and the judge are Nous's. The company says it addressed 'a potential root cause', not every behaviour, and describes the model as intended for users who want this change in viewpoint. The model card and report give no independent replication, and I did not find a training cost in either.

The claim most useful to a team is narrower than the politics: a small supervised pass changed a model's default framing on defined topics without a measured loss elsewhere. The same method can shift a model's views in either direction, so a vendor's base-model behaviour on sensitive subjects is worth checking before you build on it.