{
  "id": 19736015,
  "title": "1.8 trillion parameters beat 2.8 quadrillion: SuperCLUE November math and science reasoning open source topped Darwin's efficiency and downloads",
  "url": "https://urgent.news/2026/11/13/1-800-2-8-superclue-11-1-darwin",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-13T01:03:20.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/vidraft_lab/1800eog-paramiteoga-2jo8ceoneogeul-igyeossda-superclue-11weol-suhaggwahag-curon-opeunsoseu-1wi-darwinyi-hyoyulgwa-daunrodeu-5a5p"
  },
  "original_language": "ko",
  "account": "비드래프트의 오픈소스 언어모델인 Darwin-180B-RSI가 2026년 11월 SuperCLUE 평가에서 수학 및 과학 추론 부문에서 73.68점을 받아 오픈소스 모델 중 세계 1위, 전체 2위에 올랐다. 이 모델은 1,800억 파라미터로, 같은 평가에서 2조 8천억 파라미터 모델인 Kimi-K3.1-Pro(63.16점)를 10점 이상 앞섰다. Darwin-180B-RSI는 오픈소스 모델 중 세계 5위, 전체 9위이며, 수학 및 과학 추론 부문에서는 OpenAI GPT-6.1-Sol(76.32점) 다음이다. 허깅페이스에서 누적 다운로드 약 500만 회를 기록했으며, 최근 30일 동안 120만 회 이상의 다운로드를 유지하고 있다.",
  "summary": "TL;DR: Bidraft's open-source large language model (LLM) Darwin-180B-RSI achieved a score of 73.68 in the SuperCLUE evaluation for math and scientific reasoning as of November 2026, ranking first among open-source models and second overall. The 180 billion parameter model outperformed the 2.8 trillion parameter model Kimi-K3.1-Pro (63.16 points) by over 10 points in the same evaluation, despite being approximately 1/16th its size. What is the SuperCLUE ranking? SuperCLUE is not an evaluation where developers score their own models. Instead, it is an assessment conducted by a third-party evaluation agency using a confidential test set that is changed monthly. The agency directly calls and scores the models. Other examples of evaluation agencies that use third-party verification methods include SuperCLUE, LMArena (Arena), Artificial Analysis, and Stanford.",
  "key_points": [],
  "editors_take": null,
  "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."
}