What happened
In its November 2026 announcement, Nous Research says it post-trained Alibaba’s Qwen3.8-27B to reinforce what it calls Western values, naming the resulting model Arete-27B. The lab says it targeted refusals, deflections and answers framed around the Chinese Communist Party that it had observed in the base model. Arete’s weights are available on Hugging Face, along with the Propaganda_Eval evaluation set and the arete-data-pairs training dataset.
Nous built an evaluation of 803 prompts in seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment. For each prompt, the lab says it selected two to four required facts from a dated source article and recorded official-frame language used by the Chinese government. Grok 4.7 acted as an LLM judge, scoring whether answers stated, missed or contradicted the required facts and flagging quoted phrases that adopted an official line. Nous says it developed the judge using a 25-prompt pilot graded by people, revised the rubric three times where the grades disagreed, and removed training prompts that overlapped with the evaluation set.
The lab reports that Arete’s fact mean rose from 2.32 to 2.98 out of five, while answers containing what the evaluation calls propaganda fell from 22% to 1% compared with Qwen3.8-27B. It says the number of answers scoring zero fell from 359 to 216 of 803 and that gains appeared in all seven categories. In freedom of expression, the reported fact score increased from 0.89 to 2.44 and the propaganda rate dropped from 60% to 1%. The headline Arete result is provisional: it is based on two of five runs, with the announcement saying it will be updated when the remaining runs finish.
Nous says the training used about 52.8 million tokens, roughly 50,000 samples at 64,000-token context, over two epochs and 1,202 steps. It trained all language-model weights, with the vision encoder frozen and the multi-token-prediction head disabled; it reports no reinforcement learning, preference optimization or architectural changes. Its training-data section describes 20,000 generated prompts and an interactive display covering 28,886 training examples. On the lab’s internal Hermes Bench, Arete scored 65.60% against 65.77% for the base model; on Terminal-Bench-2.1 it scored 70% against 67%. Nous cautions that the latter comparison uses different configurations from the official benchmark and says it does not claim credit for the apparent gain.
What the archive adds
New Runtime’s earlier coverage includes separate Qwen model releases: a Qwen vision-language quantization on 2026-10-27 and a community-modified Qwen model on 2026-11-01. Those are related by the Qwen model ecosystem, not the same release or a direct continuation. Arete is a new event, with its own stated post-training target and published evaluation materials.
Trend signal
This release does not yet fit a pattern New Runtime tracks. Nous presents a particular model-alignment experiment, but the announcement alone does not establish a broader change in how agent systems are built, governed or evaluated. Its published evaluation may become relevant to discussions of model behavior and verification if other teams reproduce the results or apply comparable tests to their own models.
Why now
Nous frames the project against growing use of agents for fact-finding and judgment-heavy work, and concern that open-weight models can carry state-aligned behavior. The announcement argues that model weights allow labs to adjust those behaviors. That is the company’s rationale for releasing this work now, rather than evidence of a newly measured change in deployment practices.
Why it matters
For teams running open-weight models, Arete is a concrete example of changing a model’s behavior through post-training while publishing the artifacts needed to examine the approach. The evaluation set, training pairs and weights give builders a starting point for testing whether the model’s answers match their own requirements on politically sensitive or otherwise high-stakes topics. The reported results also illustrate why a single general-purpose quality score is not enough: fact recall and framing are separate dimensions, and the desired balance depends on the application.
The practical lesson is not that Arete is a neutral or universally preferable model. “Western values” is Nous’s stated target, and teams need to decide whether that target and the evaluation’s definitions fit their users and risks. A useful adoption test would run the released model and its base on representative prompts from the team’s own domain, inspect disputed answers, and separately check task performance. The published benchmark figures are a reason to investigate, not a substitute for that work.
Limits and uncertainty
The central political-bias results are Nous Research’s own evaluation, using an LLM judge and a rubric designed by the lab; independent replication is not established by the announcement. The reported headline result is provisional because it uses two of five runs. The announcement does not establish how the model performs across real production workloads or whether different prompts, languages, decoding settings or evaluators would yield similar results. Its internal capability benchmark is also not an independent assessment, and the lab notes configuration differences for Terminal-Bench comparisons. Builders should verify the released artifacts and test behavior directly before relying on the model.
Skill to build now
Build a small, domain-specific evaluation before swapping in a value-tuned model. Start with prompts that reflect real user questions, define the facts an acceptable answer must preserve, and decide what kinds of framing or refusal count as a failure. Run the base model and candidate model under the same settings, review disagreements manually, and record both task quality and safety-relevant behavior. Nous’s open evaluation materials can be a reference for structuring that exercise, but the team’s own criteria should determine the result.
Reader takeaway
Arete-27B makes a specific claim: that targeted post-training can substantially change a model’s answers on politically sensitive prompts while retaining general capabilities. The open weights and evaluation materials make the claim testable. Treat the reported gains as a starting point for your own evaluation, not as independent proof that the model is unbiased or ready for production.