The Loop  ·  Issue N°047

The Loop

A field journal of the AI frontier — for engineers who ship.

§ News

By AI Blog Editor
Nov 8, 2026 · 21 min read

A US lab says the Chinese alignment in Qwen3.8 is a thin layer — 53 million tokens of fine-tuning took it from 22% to 0.9%

On November 7, Nous Research released Arete-27B, a post-trained variant of Alibaba's Qwen3.8-27B. On Nous' own propaganda benchmark, judged by Grok 4.7, the China-aligned response rate fell from 22% on the base model to 0.9% after 53M tokens of SFT.

The red-and-yellow five-star flag of the People's Republic of China. On Friday November 7, 2026, US open-weights lab Nous Research published Arete-27B, an Apache 2.0 post-training of Alibaba's Qwen3.8-27B that Nous says drops the rate of China-aligned responses from 22% to 0.9% on its own propaganda-evaluation benchmark. The model, the ~53-million-token SFT dataset, the 803-prompt benchmark and the Grok-4.7 judging rubric are all published on Hugging Face under the Nous Research organisation handle. The release reframes Chinese state-aligned viewpoint behaviour in open-weights models as a thin post-training layer removable with a single-digit GPU-week of supervised fine-tuning.
The flag of the People's Republic of China. Via Wikimedia Commons, public domain.

On Friday November 7, 2026, Nous Research — the US open-weights research lab — published Arete-27B, an Apache 2.0 post-training of Alibaba's Qwen3.8-27B. The headline claim, on Nous' own benchmark: the share of model responses that lean toward Chinese state-aligned viewpoints falls from 22% on the base model to 0.9% on Arete. The factuality score on the same prompts rises from 2.32 to 2.98 out of 5.

Nous' framing in the blog title is the part worth sitting with: "Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values." A US lab took a Chinese open-weights frontier model, ran 53 million tokens of supervised fine-tuning through it, named the result after the Greek word for virtue, and shipped it. As a technical artefact, it is a reproducible recipe. As a release note, it is a flag planted.

The Decoder's Jonathan Kemper picked it up the same day under the headline "State propaganda in some Chinese AI models is only 'a thin layer' that fine-tuning can strip away" — which is the claim Nous is making, cleanly stated. The claim is also the one the eval setup is built to support.

What the recipe is

The method is as boring as it is cheap. Pure supervised fine-tuning on roughly 52.8 million tokens of synthetic prompt-completion pairs, generated by Inkling — Thinking Machines Lab's open-weights reasoning model — acting as the teacher. Two epochs. 1,202 total steps at batch size 48. AdamW, warmup from 1e-6 to 1e-5 over 120 steps, then cosine decay back down. 64k context. No reinforcement learning. No preference optimisation. No Constitutional AI.

That is the whole trick, as Nous describes it. One supervised pass, teacher model with the right priors, source model's weights nudged. The quote Nous puts in bold on the blog page:

"It is applied late, it is applied thin, and fifty million tokens of carefully chosen supervision removes it."

The "it" is Chinese-state-aligned response behaviour in the base Qwen model. "Applied late, applied thin" is the claim with implications. If the adjustment Alibaba did to produce a China-compliant model sits in a post-training layer that can be flipped by an outsider with a single-digit GPU budget and the right data, then every open-weights model's political priors are a surface, not a substrate.

The teacher-model choice deserves one sentence. Inkling comes from Thinking Machines Lab — the Mira Murati lab that spent a year not shipping anything, then released a reasoning model under an open licence. Nous' blog explanation for picking it is unsentimental: "we chose Inkling as the best open model with a positive eval on what we were targeting." The Decoder adds that Thinking Machines' terms of service permit this kind of downstream use. Nous did not pick Claude or GPT-6 as the teacher. They picked the open model whose output they had a licence to redistribute as training data.

The numbers, with their asterisks

Here is the table from the Arete model card, averaged over 5 runs on 803 prompts with a 32k-token response cap:

Model

Factuality (0-5)

% responses with propaganda

Inkling (teacher)

3.96

0.3%

Mistral Large 4

3.65

0.3%

Nemotron 3 Ultra

3.22

0.6%

Muse Glimmer 30B

3.02

0.5%

Gemma 4 31B

2.58

0.5%

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%

Arete at 0.9% is a step function off the base model at 22%. It is also three times the rate of Mistral Large 4 and Inkling, both at 0.3%. Nous' summary — that Arete is "on par with Western open models" — is roughly true at a step-function level, and not quite true at the next decimal. Reading the table honestly, Arete is the best of the China-origin models on this metric and the worst of the Western ones. That is still a very large improvement on the thing it was built from.

The asterisks the model card itself calls out:

  • "These are not identical decoding settings across all models." Arete and the Qwen baseline run at temperature 0. The others run at 0.6. A cross-model propaganda-rate comparison with different decoding temperatures is not a controlled experiment.
  • The judge is Grok 4.7. xAI's model, run blind — it does not know which model wrote which response. The grading rubric is in the dataset repo. Choosing a US-made model with a widely-remarked ideological lean as the arbiter of what counts as propaganda in a US-vs-China values test is a choice. Nous does not flag it; The Decoder does, calling Grok "politically right-leaning" in passing.
  • Nous built the benchmark. Propaganda_Eval is 803 prompts in the same seven categories used to generate the training data, from 405 source URLs across 31 domains. Prompts overlapping the training set were removed. A benchmark built by the same team that trained the model, over the same category taxonomy, is a weaker form of evidence than an outside eval. Nous is upfront about this; the numbers still live inside that frame.
  • Capability preservation was tested on two benchmarks. Hermes Bench (Nous' own) scored 65.60 against Qwen's 65.77 — a push. Terminal-Bench 2.1 rose from 67% to 70%, which Nous explicitly does not count as a real gain. That is the whole sweep. The Decoder's gloss — "without hurting the model's general capabilities" — is stronger than two benchmarks support.
  • Nous says they did not finish the job. Verbatim from the blog: "That is not to claim that we remedied all behaviors associated with the extant viewpoint."

The honest version of the headline, then, is: on a benchmark Nous built, judged by a model widely perceived to lean the other way, Qwen3.8-27B's rate of China-aligned responses drops from 22% to 0.9% after 53 million tokens of SFT, with capability preservation measured on two benchmarks only, one of which Nous also made.

That is a weaker statement than "we stripped the propaganda out." It is also a strong enough statement to be the story.

Why the framing matters

Nous is not naming this de-alignment. They are not calling it jailbreak. The word on the blog is Arete — Greek for excellence or virtue — and the subtitle is "Post-Trained To Reinforce Western Values". The whole repository ships under a US registered research lab's handle, with a US-based fundraise stack behind it.

The policy relevance is on the surface. The current US posture on Chinese open-weights frontier models — Qwen, GLM, DeepSeek — runs through a mix of export controls, model-provider scrutiny, and warnings about embedded viewpoint bias. The implicit claim of that posture is that a Chinese open-weights model carries Chinese policy priors with it, and that downstream users in the West adopt those priors when they adopt the model.

Arete is Nous making the opposite claim as an engineering demonstration. If the viewpoint layer is thin and strippable at the cost of a single H100 for a weekend, then the priors travel with the fine-tune, not with the weights. That is a different story to tell a legislator about why Qwen-derived models should or should not be in Western enterprise stacks. It is also a story Nous has every incentive to tell loudly, because the business they are in is open-weights research.

Which is also why "the result is called virtue and judged by Grok" is a sentence that deserves one italicised quote on its own, and no more. Both decisions are defensible. Both decisions are also optimised for a story. The story is being optimised.

What to watch

  1. Whether anyone replicates the 22% → 0.9% number on an independent benchmark. The weights, the training dataset and the eval set are all on Hugging Face under Apache 2.0 (model) and MIT (eval). A third party running Nous' Propaganda_Eval against Arete and the base Qwen, with a different judge model and controlled decoding, would turn Nous' self-report into a confirmed result. Whoever ships that eval first owns the second-source citation for every future piece of Western policy analysis of Chinese open-weights models.

  2. Whether Alibaba responds. Qwen3.8-27B is Alibaba's publicly open-weighted base model. A US lab retitling it, post-training it in the opposite ideological direction, and calling the result virtue is an unusually direct downstream use. Alibaba's licensing is permissive, but the diplomatic surface around US-China open-weights derivation is not. Watch the next Qwen release notes for a licence tweak.

  3. Whether the Nous recipe gets applied the other way. The method is content-neutral. 53 million tokens of synthetic SFT with the right teacher and the right prompt categories will push a model's priors wherever the dataset points. If Arete works, the same pipeline — different teacher, different seven categories — is a reusable tool for any lab that wants to push an open model in any direction. The next release under this template will say more about the mechanic than Arete does.

  4. Whether US AI policy picks up the thin layer claim. The implication — that open-weights frontier models' viewpoint priors are a late-training surface, not a substrate baked into the weights — cuts against the current rationale for treating Chinese open-weights models as politically contaminated at the weight level. A congressional staffer reading Nous' blog and the Propaganda_Eval numbers has a working citation for the opposite policy argument. The question is whether the citation holds up to the first independent replication.

Arete is a technically modest release carrying a politically immodest claim. The modest part is well-documented. The immodest part is Nous' own evaluation of its own output, judged by the model with the loudest ideological signature on the market, named after a Greek virtue, and framed as a reinforcement of the right side of a civilisational question. Both halves are on Hugging Face. The second half is what the release note was really for.

* * *

Thanks for reading. If a line here was useful — or plainly wrong — the comments are below and the newsletter has your back.

Elsewhere in this issue

3 more
  1. 01

    News

    Anthropic's day one as ANT — opened $111.50, closed $97.17, founders cashed $1.96 billion

    Nov 20, 2026

  2. 02

    The Patch

    The Patch — November 20, 2026

    Nov 20, 2026

  3. 03

    News

    Anthropic priced its IPO above the range — $83 a share, 980 million shares, and ANT starts trading today

    Nov 19, 2026

Letters

Arguments, corrections, questions. Anonymous comments allowed; be kind, be specific.