IMPLICATOR
Nous Research has released Arete-27B, a version of Alibaba's Qwen3.8-27B post-trained to reinforce Western values. On the lab's own evaluation, the share of answers containing propaganda fell from 22% for the base model to 0.9%. Nous argues the result shows that the bias layer in a Chinese-aligned model is shallow, and that the values in any open model can be changed with one fine-tune.
The evidence is the lab's own test, scored by a judge Nous developed.
What Changed
- Nous Research released Arete-27B, a version of Alibaba's Qwen3.8-27B post-trained to reinforce Western values.
- On Nous's own evaluation, answers containing propaganda fell from 22% to 0.9%, and the fact score rose from 2.32 to 2.98 out of 5.
- Training used about 52.8 million tokens of supervised fine-tuning over two epochs, with no reinforcement learning.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
What Nous measured
Nous said it observed refusals, deflections and CCP-framed answers on politically sensitive topics in Qwen3.8-27B.
The evaluation spans seven value categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment. Each prompt carries two to four required facts drawn from a dated source article. Grok 4.7 from Elon Musk's xAI served as judge. It marked each fact as stated, missed or contradicted, and it had to cite the exact phrase before scoring an answer as propaganda.
Before training, Qwen3.8-27B scored 2.32 out of 5 on facts. GLM 5.3 came in at 15% and DeepSeek V4.1 Flash at 16%.
About 53 million tokens
Nous fine-tuned Qwen3.8-27B on about 52.8 million tokens over two epochs and 1,202 steps. It used no reinforcement learning, no preference optimization and no architectural modification. Nous said it first removed every training prompt that overlapped the evaluation set.
FREE WEEKDAY MORNING BRIEFING
Track how open models get their values.
The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
About five minutes. No hype. No spam.
The answers it trained on came from a teacher model, Inkling from Thinking Machines Lab. Nous chose an open model to avoid terms-of-use restrictions on what it called "illicit distillation." It had strongly considered Nvidia's Nemotron 3 Ultra, but found evidence that it too had a "checkered record on China-related topics." Nous theorizes that Nemotron picked up the bias from working with Chinese open models during training, and says it has experienced similar contamination itself.
Thinking Machines used models including the Chinese Kimi K2.5 for Inkling's synthetic training data.
The results
After training, Arete-27B's fact score rose to 2.98 out of 5. The number of answers scoring 0 fell from 359 to 216 of 803. The improvement held in all seven categories. Freedom of expression moved the most: its fact score rose from 0.89 to 2.44, and its propaganda rate fell from 60% to 1%.
On Hermes Bench, Nous's internal general-capability benchmark, Arete scored 65.60% against 65.77% for base Qwen, a gap Nous calls indistinguishable from noise. On Terminal-Bench-2.1 it scored 70% against the base model's 67%. Nous does not take credit for that gain, citing configuration differences.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
What the numbers leave open
Every figure here comes from Nous's own evaluation. Nous developed the judge on a 25-prompt pilot it graded by hand. The Arete and base Qwen figures above come from runs at temperature 0, while the other models ran at 0.6. At 0.6, Arete scores 2.88 on facts.
Nous also stops short of a full cure. "That is not to claim that we remedied all behaviors associated with the extant viewpoint, but that we have addressed a potential root cause," the lab writes.
A recent ValsAI study found that Chinese models do not all behave the same way, and that even Western models occasionally adopted official Beijing positions.
A turn for Nous
Arete moves Nous away from Hermes 3, which in 2024 was deliberately neutral and designed to adapt to whatever worldview was set in the system prompt. Nous raised $90 million in a Series B round in October at a $1.5 billion valuation. The weights, test dataset and training data for Arete-27B are available under an Apache 2.0 license on Hugging Face.
"It is applied late, it is applied thin, and fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath," Nous writes of the bias layer. It closes with a request: "We encourage the community to run the eval on our model, on theirs, on anything they choose, and to continue experimenting with value-alignment."
Frequently Asked Questions
What is Arete-27B?
Arete-27B is a version of Alibaba's Qwen3.8-27B that Nous Research post-trained to reinforce Western values. The weights, test dataset and training data are available under an Apache 2.0 license on Hugging Face.
How much did the propaganda rate change?
On Nous's own evaluation, the share of answers containing propaganda fell from 22% for the base model to 0.9%. The largest change came in freedom of expression, where the propaganda rate fell from 60% to 1%.
Who judged the answers?
Grok 4.7 from Elon Musk's xAI served as judge. It marked each required fact as stated, missed or contradicted, and had to cite the exact phrase before scoring an answer as propaganda.
Did the model lose general ability?
On Hermes Bench, Nous's internal benchmark, Arete scored 65.60% against 65.77% for base Qwen, a gap Nous calls indistinguishable from noise. On Terminal-Bench-2.1 it scored 70% against 67%.
Which teacher model did Nous use?
Nous trained on answers from Inkling, an open model from Thinking Machines Lab. Thinking Machines used models including the Chinese Kimi K2.5 for Inkling's synthetic training data.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Free AI briefing · Weekdays
The AI stories that matter, sourced and explained.
Join the Morning Briefing. It goes out every weekday at 4:45 a.m. Pacific, with later sends for the East Coast, Berlin and Tokyo.
Free when you sign up: The Paperclip Compendium, our tested guide to running AI agents.
Free. Unsubscribe in one click.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
