Skip to content
NewA calmer, faster ZeroHour
StoryModel releaseinfo

Nous Research Releases Arete-27B Cutting Qwen3.8-27B Propaganda from 22% to 0.9%

3 sources · 3 articlesFirst reported Updated ()1
What's new

November 8: Added model sizes (53.81 GB BF16, 17.20 GB Q4KM), open release of training data and eval set, and 803-prompt evaluation details.

Merged summary · x-ai/grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

Nous Research released Arete-27B, a fine-tuned Qwen3.8-27B model that reduced propaganda from 22% to 0.9% while raising fact scores from 2.32 to 2.98.

Nous Research released Arete-27B, a 27B open-weight model obtained by supervised fine-tuning of Alibaba's Qwen3.8-27B on 52.8-53 million tokens over two epochs and 1,202 steps. The model cuts state propaganda answers from 22% to 0.9% on an 803-prompt evaluation while improving factual accuracy to 2.98/5 across seven value categories including freedom of expression. The resulting Arete-27B trended at #30 on Hugging Face, with weights released under Apache 2.0 along with training data and evaluation sets.

Key facts
  • Propaganda answers dropped from 22% to 0.9% on 803-prompt eval
  • Fact score rose from 2.32 to 2.98 across seven categories
  • Trained 1,202 steps over two epochs on 52.8M-53M tokens
  • Arete-27B weights released under Apache 2.0
  • BF16 size 53.81 GB; Q4KM quant 17.20 GB
  • Trending at #30 on Hugging Face
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

Timeline

dated events in the reporting · oldest first
  1. Nov 6, 2026Nous Research post-trains Qwen3.8-27B into Arete-27B trending on Hugging Face · Hugging Face trending models
  2. Nov 7, 2026Propaganda in Chinese AI models described as thin layer removable by fine-tuning · The Decoder
  3. Nov 8, 2026Nous Research releases Arete-27B reducing propaganda from 22% to 0.9% · MarkTechPost

What each outlet adds

beyond the shared facts
  • Hugging Face trending modelsReports #30 trending rank and 1,202 training steps with Grok 4.7 judge
  • The DecoderDescribes propaganda as shallow removable layer, with bias limited to specific categories
  • MarkTechPostDetails 53.81 GB BF16 and 17.20 GB Q4KM sizes plus open-sourced training data and eval set
Vendors
Nous ResearchAlibaba
Products
Qwen3.8-27B
Organizations
Hugging Face
AI models
Arete-27BQwen3.8-27B

Coverage

oldest first · one row per article · UTC
  1. · 2d ago Hugging Face trending modelsFirst report

    Nous Research releases Arete-27B with reduced propaganda

    Post-trained Qwen3.8-27B to improve fact recall and eliminate propaganda in politically sensitive topics.

  2. · 2d ago The Decoder

    Nous Research fine-tunes Qwen3.8-27B to strip Chinese propaganda

    Nous Research fine-tuned Alibaba's Qwen3.8-27B on 53 million tokens to drop propaganda from 22 percent to under 1 percent in Arete-27B while keeping capabilities intact.

  3. · 1d ago MarkTechPost

    Nous Research Releases Arete-27B Fine-Tuned Qwen Model

    Arete-27B is a 27B open-weight model reducing propaganda answers from 22% to 0.9% on an 803-prompt eval.