Nous fine-tunes Qwen3.8-27B to return fewer propaganda-style answers

single source· 1 articles · confidence: low · first seen 2026-11-08 04:45 UTC

What this means for you

Nothing to act on yet. The only numbers come from Nous, with no evaluation date, harness or prompt set published, so the drop from 22% to 0.9% cannot be checked, reproduced or compared against another fine-tune of the same base model.

Nous Research has published Arete-27B, an Apache 2.0 fine-tune of Qwen3.8-27B that it says answers fewer prompts in a propaganda register. The claim rests on Nous's own 803-prompt evaluation, in which the share of answers it classes as propaganda fell from 22% on the base model to 0.9% on the fine-tune. Nous gives no evaluation date, harness or prompt set, and MarkTechPost's account reports no independent reproduction. Nous says Hermes Bench, its own suite, was unchanged at 65.60%.

Models in this story

Key facts

  • ·Nous Research released Arete-27B, an Apache 2.0 fine-tune of Qwen3.8-27B. source
  • ·On Nous's 803-prompt evaluation, the share of answers classed as propaganda fell from 22% to 0.9%. source
  • ·Nous reports Hermes Bench unchanged at 65.60% after the fine-tune. source
  • ·No evaluation date or harness is given for either the 803-prompt evaluation or the Hermes Bench score. source

What the sources say

  • MarkTechPost — Short write-up of Nous's release, carrying the licence, the eval numbers and the unchanged side benchmark.

Sources

The original reporting. Follow these — they did the work.

← the wire