7 November 2026

Nous releases fine-tuned 27B model that cut propaganda output sharply

• Nous Research fine-tuned a 27-billion-parameter open model, Qwen3.8-27B, to reduce propaganda in its answers.
• On an 803-prompt test set, the share of propaganda-style replies fell from 22% to 1%.
• The team reports no significant loss in general performance from the change.

How it was covered

Latent Spaceswyx & Alessio

Fine-tuning data size, values eval results, Hermes Bench regression check, released artifacts