Summary
According to a post on X by @NousResearch, Nous Research has released Arete-27B, a fine-tune of Qwen3.8-27B aimed at Western values. The lab says that on an 803-prompt evaluation covering freedom of expression, democracy, privacy and related categories, the model's propaganda rate fell from 22% to 1% after about 53M tokens of post-training. It also reports no significant drop on Hermes Bench. The evaluation, weights and data are publicly available, according to the post.
AI-assisted summary based on the listed source.
What happened
Nous Research releases Arete-27B, a Qwen3.8-27B fine-tune aimed at Western values. On an 803-prompt eval across freedom of expression, democracy, privacy, and related categories, the lab says the model’s “propaganda rate” fell from 22% to 1% after ~53M tokens of post-training, with no significant Hermes Bench...
Why it matters
Developers and researchers working with open-weight models can inspect the published evaluation, weights and data directly rather than relying only on the lab's claims. The headline frames the release as removing China-aligned defaults, a framing that depends on how the evaluation defines propaganda, so the methodology is worth checking. Watch for independent replication of the 22% to 1% result.
Signal Intelligence
Signal Strength 90%
Technical label WATCH
Public Interest 58
Category PRIVACY
Reader Depth GENERAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 73
Practical Impact Score 26
Novelty Interest Score 94
Consequence Score 34
Curiosity Score 36
Shareability Score 66
Why this is here
VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).
