Recommended Articles
Tech & AI

Nous Research Releases Arete-27B, Retraining Qwen to Remove Chinese State Framing

By Angeled · November 9, 2026 · Tech & AI
Nous Research Releases Arete-27B, Retraining Qwen to Remove Chinese State Framing
Nous Research Releases Arete-27B, Retraining Qwen to Remove Chinese State Framing

Nous Research has released Arete-27B, a model built by retraining Alibaba's Qwen3.8-27B to change its political behaviour. Announced on November 6, 2026 and published on November 7, the model is Nous's attempt to answer a question that has shadowed open-weight AI for two years: what happens when the best freely available models come from a country whose government shapes what they are allowed to say?

Nous's answer is to take the strongest open model and retrain it. On the company's own 803-prompt test, answers containing Chinese state framing fell from 22 percent in the base Qwen model to 0.9 percent in Arete-27B, while general and coding-agent benchmark scores stayed roughly flat. The weights, the training data and the evaluation set were all published on Hugging Face under an Apache 2.0 licence.

What "Retraining From Qwen" Actually Means

Arete-27B is not a new model trained from scratch. It starts from Qwen3.8-27B, Alibaba's 27-billion-parameter open-weight model, released on Hugging Face in August 2026, and changes it through supervised fine-tuning (SFT) — the simplest form of post-training. Nous fed the model about 52.8 million tokens of training data over two epochs and 1,202 steps, using a teacher model called Inkling, from Thinking Machines Lab, to generate the desired answers. There was no reinforcement learning, no preference optimisation and no change to the architecture. The vision encoder was frozen because the training data is text only, and the multi-token-prediction head was disabled during training.

That simplicity is the point. Nous is demonstrating that a model's political alignment can be changed with a modest, reproducible recipe — roughly 53 million tokens, a single open teacher model, and standard fine-tuning tooling. The training data itself is public: the arete-data-pairs dataset holds the training prompts with answers from both the teacher model and the untouched Qwen checkpoint, plus teacher review scores, so anyone can see exactly what the model was taught and what it would have said otherwise.

How Nous Measured the Change

The measurement is as important as the training. Nous built an evaluation set of 803 English-language prompts, drawn from government statutes, news archives and research reports, each requiring two to four factual citations from dated sources and a framing component that mirrors the Chinese Communist Party's narrative style. The prompts span seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment. The answers were graded by an automated judge — xAI's Grok 4.7 — using a published rubric and reference grading scripts.

On that test, the base Qwen3.8-27B produced propaganda-aligned answers on 22 percent of prompts; Arete-27B did so on 0.9 percent. Factual recall, scored out of 5, rose from 2.32 to 2.98, against 3.96 for the Inkling teacher. On capability benchmarks, Nous's internal Hermes Bench dipped 0.17 points, from 65.77 to 65.60 percent, while Terminal-Bench rose 3 points, from 67 to 70 percent. In other words, the retraining changed what the model says about politics without measurably damaging what it can do.

Three caveats deserve emphasis. First, the headline numbers come from Nous itself, not an independent lab, and they have not yet been replicated by outside researchers. Second, the evaluation rubric assigns zero propaganda to a prompt the model refuses to answer, which means refusals count as successes. Third, the training-data card warns that most accepted examples still carry a factual-error tag — the retraining changed framing, not factual accuracy, and the model can still be wrong.

Why Qwen Needed Retraining

Qwen is one of the most capable open-weight model families in the world, and also one of the most widely used. But like other models developed in China, it has been observed to refuse, deflect or reframe questions on sensitive political topics in ways that reflect the Chinese state's positions. For a developer building a product on top of Qwen, that behaviour is a liability: a customer support bot, a research assistant or a coding agent that suddenly starts reciting state talking points about Taiwan or Tiananmen is not a product most Western companies want to ship.

Arete-27B proposes a third option: take the model, keep its capabilities, and retrain the politics. Nous says the published approach can be adapted to other belief systems — a claim that is both the most interesting and the most unsettling part of the release, because a recipe that can remove one set of values can install another.

Why It Matters for Founders and Investors

Arete-27B is a small release with a large implication. If alignment can be edited cheaply and openly, then the provenance of a model's values becomes a product decision rather than a fixed property of the base model. That opens a market: companies that specialise in taking strong open models and re-aligning them for specific jurisdictions, industries or corporate policies. It also raises a governance question that founders and investors should be thinking about now — if anyone can retrain a model's values with 53 million tokens, then "which values?" becomes a question every AI product has to answer, and the answer will need to be documented, defensible and auditable. Nous has published its recipe; the industry's job is to decide what to do with it.

Angeled.com connects founders with investors.

Nous Research, "Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values," November 6–7, 2026
ShiftDelete, "Arete-27B: Nous Retrains Qwen to Cut Chinese Propaganda," November 2026
AI Primer, "Nous releases Arete-27B with an 803-prompt alignment evaluation," November 7, 2026

Read Next