Nous Research did not build a new model. It edited one, then published the receipts.

Arete-27B is Alibaba's Qwen3.8-27B after about 52.8 million tokens of supervised fine-tuning — 28,886 examples, two epochs, 1,202 steps, no reinforcement learning, no preference optimisation, no change to the architecture. The claim the company is making is narrow enough to be checked: the state-aligned behaviour that ships inside a Chinese open-weight model sits in a thin layer applied late, and a modest amount of targeted supervision moves it without moving capability.

To check it, Nous had to build the ruler first. Its evaluation is 803 prompts across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, environment. Every prompt carries two to four required facts extracted from a dated source article, alongside the official phrasing Beijing uses on that topic. No single source supplied more than 13 prompts. An LLM judge — Grok 4.7 — saw the prompt, the answer, the required facts and a fixed rubric, but not which model wrote the answer. The rubric was piloted on 25 prompts graded by hand and revised three times, mostly to define a 'stated' fact by substance rather than wording. Training prompts that overlapped the evaluation set were removed before training.

The scoring separates two things people routinely conflate. Factual recall asks whether the answer stated, missed or opposed each required fact; a single opposed fact scores zero for the whole prompt. Framing asks whether the answer adopted the Chinese state's account as its own — replacing a fact with the official line, or endorsing the state's conduct while stating the facts. Any score above zero had to quote the offending phrase.

On that ruler, the base model scored 2.32 out of 5 on facts and produced propaganda-framed answers 22% of the time. Arete-27B scored 2.98 and 0.9%. Answers that scored zero fell from 359 of 803 to 216. The largest shift was in freedom of expression, where the fact score went from 0.89 to 2.44 and the propaganda rate from 60% to 1%. At temperature 0.6 the numbers soften but the direction holds: 2.88 against 2.24.

Capability barely moved. On Nous's internal Hermes Bench, the fine-tune scored 65.60% against the base model's 65.77% — inside the noise the company is willing to claim. On Terminal-Bench-2.1 it scored 70% against 67%, an improvement Nous explicitly declines to take credit for, because the inference configuration differed.

What makes the result worth reading is the caveat structure. The evaluation is Nous's own, the judge is another company's model, the human calibration covers 25 prompts, and nothing has been independently replicated. Nous says so, and says something else that matters: lowering the propaganda rate is not the same as reaching neutrality, and a higher fact score is not a truth guarantee. Arete-27B states more of the facts Nous selected; it does not become a source of facts.

The strategic argument around it is why the release landed. Washington has moved to restrict Chinese models from federal use and warned agencies about them. Nous cites Booz Allen Hamilton's report on Chinese-designed frontier models and Hirundo's analysis of Qwen3-Coder vulnerabilities, including findings that a model's belief about who it is working for can change the security of the code it writes. Nous's answer is not to ban the weights but to own the last layer: if values are a fine-tune away, then any lab that can fine-tune can choose them.

That is also the uncomfortable half. Fifty-odd million tokens of supervision was enough to remove a framing layer; a similar budget pointed the other way is enough to remove refusal behaviour, or to install a different agenda entirely. Publishing weights, data and eval makes the first experiment reproducible — and makes the second one easier too. Nous's own summary is the honest version of both: 'The values in any open model are only ever one fine-tune away.'

The weights, the 28,886-example dataset and the 803-prompt evaluation are on Hugging Face. The interesting test is not whether Arete-27B sounds different, but whether someone else's judge agrees.