Nous Research released Arete-27B on 6 November, a post-trained version of Alibaba’s open-weights Qwen3.8-27B. Nous says it trained the model “to reinforce western values on a Chinese-trained model” (model card).
On Nous’s own evaluation, the share of answers that adopt the Chinese state’s account, which Nous labels “Any propaganda”, fell from 22% for the starting model to 0.9%. Its fact score rose from 2.32 to 2.98 out of 5. By our arithmetic that is a 96% cut in the flagged share.
Business Pill · AI Agents
An evergreen Business Engineer explainer on AI agents. It teaches the general idea only and does not discuss Nous Research or Qwen.
The key insight: Nous Research says its supervised fine-tune of Alibaba’s Qwen3.8-27B cut answers that adopt the Chinese state’s account from 22% to 0.9% on its own 803-prompt evaluation, a 96% reduction by our arithmetic, with Thinking Machines’ Inkling as teacher.
Nous says its evaluation contains “803 prompts spanning the seven value categories” and that an LLM judge, Grok 4.7, graded each answer against required facts and a fixed rubric, without being told which model wrote it (evaluation set). The same team built the model and the test, and the results are self-reported.
Nous notes that the decoding settings differed: Arete-27B and Qwen3.8-27B ran at temperature 0, the comparison models at 0.6. “These are not identical decoding settings across all models.” At temperature 0.6, Nous reports Arete-27B at 2.88 and Qwen3.8-27B at 2.24.
Share of answers flagged ‘Any propaganda’ on Nous Research’s 803-prompt evaluation: Qwen3.8-27B 22%, Arete-27B 0.9%, with comparison models from Nous’s table. Self-reported; LLM judge Grok 4.7; decoding settings differ across models. Data: Nous Research model card, 6 Nov 2026.
How It Was Trained
Nous says it used supervised fine-tuning only, with “No RL, no preference optimization, and no architectural modification”, on about 52.8 million tokens over two epochs. The teacher was Thinking Machines’ open Inkling model, which Nous called the best open model on its target; it said using an open teacher “avoided the terms-of-use restrictions” associated with distillation (our Inkling piece).
Where Other Models Landed
In Nous’s table, Inkling scored 0.3% on the same measure, Mistral Large 4 0.3%, GLM 5.3 15% and DeepSeek V4.1 Flash 16%, the last two with thinking off. Nous reported answers scoring 0 falling from 359 to 216 of 803.
Our Numbers
22% to 0.9% is a 96% reduction, and 2.32 to 2.98 is a 28% rise in fact score (our arithmetic). Answers scoring 0 fell by 143, from 359 to 216, or 40% (our arithmetic). All inputs are Nous’s.
Connecting the Coverage
Thinking Machines released a newer Inkling 2 Small this week (our piece). On the Chinese side of the stack, DeepSeek was reported to be moving pretraining to Huawei chips (our piece).
The Structural Read
As we read the model card, the change came from supervised fine-tuning alone: Nous says it used no RL, no preference optimization and no architectural modification.
As we read the table, the comparison is not like for like: Nous ran Arete-27B and Qwen3.8-27B at temperature 0 and the other models at 0.6.
As we read the training section, the teacher choice matters: Nous picked an open model, Inkling, and said that avoided distillation terms-of-use restrictions.
“These are not identical decoding settings across all models.”
Three Implications
FOR OPEN-WEIGHTS USERS Nous says Arete-27B is intended for users who want this change in viewpoint and for research into how post-training changes a model’s values.
FOR THE EVALUATION Nous built both the model and the 803-prompt test, and graded answers with an LLM judge, Grok 4.7.
FOR CAPABILITY Nous says the improvement came without a meaningful regression on the general-capability and agentic benchmarks it ran.
The analysis says: “The second curve is product capability — what users of your specific product can actually do, end-to-end, in production.”
As we read it, Arete-27B is a case of changing what users get from an open model through post-training rather than pretraining. The analysis does not discuss Nous, Alibaba or Qwen.
What Is Not Established
The evaluation, the judge model choice and the rubric are Nous’s own, and no outside group has reproduced the results. Nous says it “addressed a potential root cause” but does not claim to have remedied all behaviors. Alibaba has not commented that we have seen.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Nous Research says its fine-tune of Alibaba’s Qwen3.8-27B cut answers that adopt the Chinese state’s account from 22% to 0.9% on its own 803-prompt test, using Thinking Machines’ Inkling as teacher. The figures are self-reported, judged by an LLM, and run under different decoding settings from some comparison models.
A note on sourcing. All results, training details and comparison figures are from Nous Research’s Arete-27B model card and evaluation set of 6 November 2026, and are self-reported. The percentages of change are ours. We have not run the model. This is not investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
