{
  "id": 18879333,
  "title": "Nous Research unveils Arete-27B, a post-trained Qwen3.8-27B model that reinforces Western values, saying the underlying Chinese model's bias layer is shallow (Nous Research)",
  "url": "https://urgent.news/2026/11/08/nous-research-unveils-arete-27b-a-post-trained-qwen3-8-27b-model-that",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-08T06:20:00.000Z",
  "source": {
    "name": "Techmeme",
    "slug": "techmeme",
    "url": "https://nousresearch.com/arete"
  },
  "original_language": "en",
  "account": "Nous Research has unveiled Arete-27B, a post-trained version of Alibaba's Qwen3.8-27B model aimed at reinforcing Western values. The new model derives its name from the ancient Greek concept of excellence and virtue, as well as its physical geography namesake—an arête, a sharp ridge of rock separating glacial valleys. The creation of open-source models allows labs to tailor intelligence to various preferences. However, a growing concern is the homogeneity of viewpoints across AI systems, with many frontier labs sharing a similar worldview and exhibiting political bias. This bias can have significant implications, particularly as users rely on AI agents for fact-finding and judgment-heavy tasks. Open-weight models help mitigate this issue, provided companies like Nous Research can operate on these weights. While open-weight Chinese models are among the strongest available, they come with behavior that aligns with state-aligned preferences. The US government has taken steps to restrict Chinese AI models from federal use and issue warnings about broader risks. The research team observed Qwen3.8-27B refusing, deflecting, and providing CCP-framed answers on politically sensitive topics. They post-trained the model to alter these biases, addressing a potential root cause rather than claiming to remedy all associated behaviors. The team built an evaluation of 803 prompts spanning seven value categories, measuring factual recall and framing. They developed a system to assess an LLM judge's ability to determine whether a model's answer stated, missed, or contradicted required facts, as well as whether it adopted the official Chinese line. The model's fact mean score and propaganda score are provided, with the evaluation available for others to replicate and compare results.",
  "summary": "We post-trained over Alibaba's Qwen3.8-27B to reinforce Western values on a Chinese-trained model. We call this new model Arete-27B.",
  "key_points": [
    "Nous Research unveils Arete-27B, a post-trained Qwen3.8-27B model",
    "Model reinforces Western values, addresses Chinese model bias",
    "Evaluation assesses factual recall and framing across 803 prompts"
  ],
  "editors_take": "Post-training open-weight Chinese models to align with different cultural preferences allows developers to counter state-aligned biases and offers users more diverse viewpoints in high-stakes decision-making tasks.",
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}