What changed
Nous Research released Apache-2.0 Arete-27B, a supervised fine-tune of Qwen3.8-27B, together with its training pairs and an 803-prompt political-framing evaluation. Nous reports that its own 'any propaganda' rate fell from 22% to 0.9% without a meaningful regression on the capability tests it ran.
Why it matters
The release makes a politically motivated post-training intervention auditable at the weights, data and evaluation layers. It also shows why values claims need separate factuality, capability and distribution-shift tests instead of a single alignment score.
What to do
Download the exact checkpoint and public datasets, rerun the evaluation with an independent judge and matched decoding, then add privately held factuality, refusal, safety and agent tasks before considering the model for sensitive assistance.
Caveat
The headline results are Nous-authored and not independently reproduced. The rubric encodes the publisher's chosen 'Western values,' uses an LLM judge, compares some models under different temperatures, and freezes the vision encoder during text-only training.