华为官宣开源盘古 openPangu-2.0-Pro 模型及技术报告。[1]该模型参数规模约 505B，每 token 激活参数规模约 18B，基于昇腾 NPU 训练，支持 512k 上下文长度。[1]训练数据总量约 34T tokens。[1]模型在 Attention 架构、拓扑架构、自投机模块等方面进行了全面升级。[1]
余承东在 6 月 HDC 2026 主题演讲中宣布将从 6 月 30 日起陆续开源 7 大组件。[1]华为已于 6 月 30 日开源了 92B 参数的 openPangu-2.0-Flash 模型。[1]此次 openPangu-2.0-Pro 的发布是华为盘古开源计划的又一组件。[1]模型权重、基础推理代码及技术报告已同步发布。[1]
Huawei has officially released the Pangu openPangu-2.0-Pro model along with its technical report as open source.[1] The model features approximately 505 billion parameters, with an activated parameter scale of roughly 18 billion per token, and was trained on Huawei's Ascend NPU architecture.[1] It supports a context length of 512,000 tokens and was trained on a total of approximately 34 trillion tokens of data.[1]
This release follows Yu Chengdong's announcement at the HDC 2026 conference in June, where he committed to gradually open-sourcing seven major components beginning June 30.[1] The company had already released the 92-billion-parameter openPangu-2.0-Flash model on June 30 as part of this initiative.[1] The Pro model represents a comprehensive upgrade across multiple technical dimensions, including improvements to the Attention architecture, topology architecture, and self-speculative modules.[1] The open-source model, weights, foundational inference code, and technical documentation are now available at https://ai.gitcode.com/ascend-tribe/openPangu-2.0-Pro.[1]
