Nous Research's Arete-27B Cuts Chinese State-Aligned Responses by 95%

Nous Research has released Arete-27B , a fine-tune of Alibaba’s Qwen3.8-27B designed to reduce answers that the lab’s rubric classifies as aligned with the Chinese government and Communist Party. Nous reports that the rate on its 803-prompt values evaluation fell from 22% for the base model to 1% for Arete. Its internal agentic benchmark found no significant capability loss, and the release includes the weights, evaluation prompts, and roughly 53 million tokens of training data.

Qwen’s defaults travel downstream

Qwen3.8-27B is a dense 27-billion-parameter vision-language model with a native 262,144-token context window and support for agentic workloads. Its combination of long context, image understanding, tool use, and downloadable weights makes it a practical foundation for products that need local deployment or model customization.

Applications built from open weights can inherit the base model’s tendencies on political, historical, and civil-liberties questions. System prompts and output filters can shape responses; recurring behavior in the weights may surface across paraphrases, languages, multi-turn conversations, or adversarial prompts. For developers, those defaults can conflict with product policies, customer requirements, or regional regulations.

Nous measured that behavior with 803 prompts across seven value categories, including freedom of expression, democracy, and privacy. Its rubric labels outputs according to whether they match positions associated with Chinese state propaganda. The reported percentages summarize that test set and scoring method, so deployments with different prompts, languages, and evaluators may produce different results.

Fifty-three million tokens, 21 points

Post-training updates an existing model using a curated set of examples, requiring far less computation than training the base model from scratch. Nous applied roughly 53 million tokens of targeted data to Qwen3.8-27B. On the same values evaluation, the state-aligned response rate fell by 21 percentage points, a relative reduction of about 95%. Nous says the resulting 1% rate falls within the range of the Western open-weight models it tested.

The training data emphasizes values that Nous describes as Western liberal principles, including free expression, democratic processes, and individual privacy. This framing matters when interpreting the result: the evaluation measures conformance to a defined value policy rather than a viewpoint-free form of alignment.

Nous checked general capability with Hermes Bench , its internal suite of 150 tasks covering research, tool use, memory, and other agentic skills. The lab reports no significant difference between Arete and the base model. Independent replication has yet to be reported, and production workloads may expose regressions that the internal suite does not cover.

An auditable release

• Model weights: Arete-27B, derived from Qwen3.8-27B.
• Values evaluation: 803 prompts spanning seven categories.
• Training corpus: Approximately 53 million tokens used for post-training.
• Capability benchmark: Results from the 150-task Hermes Bench suite.
• Methodology: Nous’s full release page .

Publishing the evaluation and training data allows researchers to inspect category definitions, examine prompt coverage, reproduce the baseline, and test other models under the same rubric. Teams can also adapt the process to a different value policy while preserving a clear record of the behavior they intended to change.

Checks before production

1. Reproduce the baseline. Record the exact model revision, chat template, system prompt, sampling settings, and evaluator before comparing results.
2. Audit the labels. Document how responses receive a state-aligned classification, including any judge model, human review process, ambiguity rules, and scoring threshold.
3. Expand the prompt set. Test supported languages, paraphrases, indirect questions, multi-turn conversations, jailbreaks, and topics outside the seven published categories.
4. Run product regressions. Evaluate tool calls, retrieval, structured output, latency, memory, safety controls, and domain-specific accuracy on representative workloads.
5. Review deployment terms. Verify the licenses for the base weights, fine-tune, evaluation set, and training corpus, along with any data-governance requirements.

One model, one rubric

Within Nous’s test setup, Arete shows that targeted post-training can sharply alter measured political behavior in a 27-billion-parameter model with no significant regression on one internal capability suite. The evidence covers one base model, one training mixture, one values rubric, and one agentic benchmark. Broader claims depend on independent evaluation across models, languages, and real applications.

Arete extends the steerability work behind Nous’s earlier Hermes releases into an explicitly defined political value set. For open-weight developers, the reusable method is straightforward: define the target behavior, publish a measurable rubric, apply targeted post-training, rerun capability tests, and release enough material for others to verify the result.