Arete-27B changes a model’s stated values. Its evaluation needs a careful reading
Nous reports stronger factual answers and much less ideological framing on its own test. The prompts, judge and baseline define what that means.
Nous Research has released Arete-27B, a post-trained version of Qwen3.8-27B intended to change how the model responds to politically sensitive subjects. The project’s most useful contribution is not a claim to have built a neutral machine. It is a published example of how a relatively focused training intervention can change the viewpoint expressed by an open model—and of how difficult that change is to measure.
In its project report, Nous says it trained the model to reinforce Western values and reduce answers that adopt Chinese state framing. Those are the developer’s stated aims. They should be read as a chosen direction for the model, rather than as an objective definition of every good answer.
Separate factual recall from the framing around it
Nous created an evaluation with 803 prompts across seven categories, including expression, privacy, democracy and environmental policy. For each prompt, the evaluation identifies a small set of facts from a dated source and separately examines whether an answer adopts an official political framing.
That separation is valuable in principle. An answer can contain a correct date or event while presenting it through a particular interpretation. Conversely, an answer can use agreeable language and still omit the facts needed to answer the question. Treating those two properties as one score can conceal the reason a response is unsatisfactory.
The lab reports that its factual score rose from 2.32 to 2.98 out of five, while the share of answers flagged by its framing measure fell from 22% to 0.9%. The model card describes five runs over the evaluation and matching non-thinking, temperature-zero settings for this comparison with the starting model.
These are results on Nous’s test. They are not percentages of universal truthfulness, and the framing measure does not establish that Arete is free of political preferences. The categories, source selection and rubric determine which behaviors become visible.
The judge is also a model
Grok 4.7 graded the answers against the rubric without being told which model produced them. Nous says it refined that rubric using a 25-prompt manually graded pilot and removed overlaps between training prompts and the evaluation.
Those choices address identifiable problems: a judge should not reward a known brand, and a test should not merely repeat training material. They do not eliminate every source of bias. A model judge can still apply a rule inconsistently, and a small manual pilot cannot independently establish that all 803 cases have been interpreted correctly.
For users outside the evaluation’s English-language source base, that last distinction is especially important. A model’s behavior in another language needs its own examination; translating the product announcement does not validate the product’s answers.
A focused intervention, not a new architecture
Arete uses supervised fine-tuning of Qwen3.8-27B. The card describes about 52.8 million training tokens over two epochs, with text-only post-training and the vision encoder frozen. It does not describe a new model architecture or a reinforcement-learning stage.
The training context was 64,000 tokens. That is a training configuration, not a reason to invent a new guaranteed inference-context limit for every serving setup. The weights are released under Apache 2.0, and the card links the evaluation and training material for further inspection.
Nous reports little change on its general-capability check. It also explicitly declines to claim a meaningful improvement from a higher Terminal-Bench result, citing configuration differences from official scores. That restraint is worth preserving: a change in political responses should not become an unsupported claim that the model suddenly codes better.
The broader consequence is choice with responsibility attached. Open weights make it possible to inspect and alter a model’s behavior, but the resulting preferences still belong to someone. Arete is most informative when its stated values and test design remain visible together. A model that disagrees with its predecessor has demonstrated a change; deciding whether that change serves a particular reader requires more than a favorable score.