华为开源openPangu-2.0-Pro:505B参数、512K上下文,全栈开源

After Vowing to Make Pangu World No.1, Yu Chengdong Opens openPangu-2.0-Pro: 505B Parameters, 180B Sparse Activation, 512K Context, First Frontier Model Fully Trained on Ascend NPU

精选理由

华为把500B级大模型开源了,权重代码报告都给了,还是在昇腾上训的,不用N卡也能跑。

AI 摘要

华为开源openPangu-2.0-Pro,发布模型权重、推理代码和技术报告。该模型拥有5050亿参数,推理时稀疏激活1800亿参数,上下文窗口达到512K。华为称它是首个完全在昇腾NPU上训练的前沿模型,也是首个完全不依赖NVIDIA硬件的500B以上开源模型。余承东此前曾表示要让盘古成为世界第一。

原文 · Pandaily

After Vowing to Make Pangu World No.1, Yu Chengdong Opens openPangu-2.0-Pro: 505B Parameters, 180B Sparse Activation, 512K Context, First Frontier Model Fully Trained on Ascend NPU

Huawei open-sources openPangu-2.0-Pro with weights, inference code, and technical report, the first 500B-plus model trained entirely on non-NVIDIA hardware, as full-stack open source strengthens domestic compute ecosystem.

  • Simon Willison’s Weblog08-02 04:16原文
  • marktechpost08-02 06:21原文
  • IT之家08-03 01:56原文
  • NVIDIA AI08-03 16:07原文