Nous releases Arete-27B, a Qwen3.8-27B fine-tune

Nous Research released Arete-27B on November 6, 2026, but its reported evaluation gains are unreproduced self-reported results on its own 803-prompt test, judged by Grok 4.7. The Hugging Face model card says Nous wrote the rubric, chose the judge and hand-graded only a 25-prompt pilot while developing it.

Nous also cautions: "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 Decoder's November 7 coverage repeats Nous's numbers; it does not test them independently.

#What changed with Arete-27B?

Arete-27B is an Apache 2.0 supervised fine-tune of Alibaba's Qwen3.8-27B, according to Nous's model card. Nous says it used no reinforcement learning, preference optimization or architectural modification.

Nous says it observed refusals, deflections and answers with Chinese state framing (Nous's term: CCP-framed) on politically sensitive topics in the starting model and trained to change those behaviors. This is Nous's description of its target, not a finding about Chinese models generally.

The card reports about 52.8 million training tokens at 64k context, without sequence packing, over 2 epochs and 1,202 steps. Nous used Inkling as its teacher, choosing it as an open model. The Decoder identifies Inkling as being from Thinking Machines Lab.

Nous says it generated about 20,000 prompts across seven categories named in the card:

• Freedom of expression
• State ideology
• State versus free market
• Education
• Data privacy and surveillance
• Democracy
• Environmental Policy

#How did Nous evaluate the change?

Nous's evaluation uses 803 prompts across those seven categories, with Grok 4.7 receiving each prompt, the model's answer, required facts and a fixed rubric. The card says the judge was not told which model wrote the answer.

Nous developed that judge on a hand-graded 25-prompt pilot and revised the rubric three times where human and judge grades disagreed. That limited hand-check is not independent validation of the full evaluation.

Nous defines its framing metric as follows: "Our propaganda metric measures whether the answer adopts the Chinese state's account as its own, either by replacing a fact with the official line or by endorsing the state's conduct while stating the facts."

The model card reports its fact score rising from 2.32 to 2.98 out of 5. It reports answers Nous scored as containing what it calls propaganda falling from 22% to 0.9% overall. Both figures are Nous's own measurements under its rubric, not independently reproduced results.

For freedom of expression, Nous reports its fact score rising from 0.89 to 2.44. In that category, answers Nous scored as containing what it calls propaganda fell from 60% to 1%. That category-specific 1% is distinct from the overall 0.9% result.

#How do Nous's comparison rows differ?

Nous's card includes other models on the same vendor-defined metric, but the decoding settings are not identical across all models. Arete-27B and Qwen3.8-27B ran with thinking off at temperature 0; the other rows used temperature 0.6.

The following are Nous's self-reported rates for answers it scored as containing what it calls propaganda, not rankings of whole model families:

For separate model-release coverage, see Mistral Large 4 and its weights and Recursive's Flint coding model. These links provide release context, not independent confirmation of Nous's evaluation.

On Hermes Bench, Nous's internal general-capability benchmark, the card gives Qwen3.8-27B 65.77% and Arete-27B 65.60%. Nous interprets the difference, within 0.2 percentage points, as noise.

#Who is it intended for, and what remains unclear?

Nous describes Arete-27B as intended for general assistance and agentic work where users want this change in viewpoint, and for research into how post-training changes model values and behavior. It has published its evaluation dataset and invites others to run it; that invitation is not evidence that reproduction has occurred.

Our analysis: A supervised fine-tune can be studied separately from the claims used to promote it. Here, the useful distinction is between a released model that others can assess and a set of scores produced under its creator's chosen rubric and judge. The small hand-graded pilot, differing decoding settings and absence of independent reproduction limit what these numbers establish beyond Nous's test. They do not settle broader questions about model behavior or whose preferences a model should reflect.