A model can recall a fact and still frame it in a way your readers would dispute.
Nous Research released Arete-27B on November 6. The model is a supervised fine-tune of Alibaba’s Qwen3.8-27B. Nous released the weights under Apache 2.0 and documented how it evaluated responses to politically sensitive questions in English.
The lab scores factual recall and framing separately
Nous defined required facts for each of its 803 prompts. A model acting as the grader checked those facts and applied the lab’s framing rubric. Under that rubric, Nous reports a drop from 22% to 0.9% in responses it classified as propaganda. Its average factual-recall score rose from 2.32 to 2.98 out of 5.
Those are lab results, not an independent audit of neutrality. The questions cover a defined set of topics, and the grading rules encode the lab’s criteria. The model card also notes different decoding settings for some other models in the comparison. Read the head-to-head result with Qwen under those test conditions.
Check what a lower framing score can hide
If you are choosing a model for research, write a few questions absent from the published set. Include one where the evidence is strong and one where credible sources disagree. Have a reviewer who does not know which model wrote each answer score citations, factual omissions, and how the answer represents disagreements. Keep those measures separate.
An answer can avoid a specified framing by dropping context the reader still needs. That is why I would look at omissions alongside the headline score. We have not run a separate comparison. The model card explains the rubric and links to the released evaluation.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
