---
title: "PanGu (Huawei) — 华为盘古"
source: "https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4"
wiki: "systems-analysis.info/int"
article: "PanGu_(Huawei)_—_华为盘古"
language: "zh"
categories:
  - "Category:Chinese"
  - "Category:Large language models"
  - "Category:LLM families"
  - "Category:Machine learning"
revision_id: 5460
wiki_created_at: 2026-09-06T23:49:37Z
wiki_modified_at: 2026-09-06T23:49:37Z
downloaded_at: 2026-09-07T23:08:36Z
---

# PanGu (Huawei) — 华为盘古

**Huawei PanGu**（中文：*盘古*）是由华为云（Huawei Cloud）开发的一系列超大规模人工智能预训练模型（基础模型）。“盘古”之名源于中国神话中的创世神祇盘古<sup>[\[1\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-hdc2021-1)</sup>。盘古模型系列涵盖了自然语言处理（NLP）、计算机视觉（CV）、多模态分析、预测建模和科学计算等多个领域。

## 历史与发展

### PanGu-α (2021) - 盘古-α (2021)

该系列的首个模型 **PanGu-α**（盘古-Alpha）于2021年4月发布。它拥有 **2000亿** 个参数，在当时成为全球最大的中文语言模型，规模超过了 OpenAI 的 GPT-3（1750亿参数）<sup>[\[2\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-aibusiness_2021-2)</sup>。

该模型由华为云团队与**诺亚方舟实验室**（Noah's Ark Lab）合作开发，并在由 **2048** 个专用**华为昇腾910**（Huawei Ascend 910）处理器组成的集群上，使用 MindSpore 框架进行训练<sup>[\[3\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-tech_report_alpha-3)</sup>。训练数据集包含了 1.1 TB 的高质量中文文本数据。PanGu-α 在**CLUE**（中文语言理解评测基准）上表现出色，取得了总榜第一的成绩<sup>[\[1\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-hdc2021-1)</sup>。

### PanGu 3.0 (2023): 平台化方法

2023年7月，华为发布了**盘古3.0**（PanGu 3.0）平台，标志着其从单一模型向面向行业应用的多层**“5+N+X”**架构转型<sup>[\[4\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-pangu3_launch-4)</sup>。

- **L0（基础层）**: 五个基础模型（NLP、CV、多模态、预测和科学计算）。
- **L1（行业层）**: N个行业模型，在基础模型之上针对特定行业（如政务、金融、制造等）进行微调。
- **L2（场景层）**: X个面向具体应用场景的模型（如虚拟助手、台风路径预测等）。

这种分层架构使客户既可以直接使用现成的解决方案，也可以在自有数据上对行业模型进行微调，从而极大地简化并降低了适配成本。

### PanGu 5.5 (2025): 混合专家（MoE）架构

2025年6月，华为宣布升级至**盘古5.5**（PanGu 5.5），旨在深入解决行业问题。其核心特点是采用了拥有256个专家子网络的**混合专家（Mixture-of-Experts, MoE）**架构，使得总参数量达到**7180亿**<sup>[\[5\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-pangu5_techmag-5)</sup>。MoE 架构允许在处理特定任务时动态激活模型的一部分，据华为称，这使得推理效率比前几代产品提高了**八倍**<sup>[\[6\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-aimagazine_5_5-6)</sup>。

## 关键架构与技术方案

盘古模型基于GPT类的 Transformer 架构构建，并进行了一系列创新以支持超大规模模型的训练。为了控制生成过程，模型引入了一个专门的**查询层**（Query Layer），有助于在预训练阶段引导模型产生期望的输出<sup>[\[3\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-tech_report_alpha-3)</sup>。

盘古模型的训练和运行与华为自有的硬件和软件平台紧密集成：

- **昇腾910处理器**（Ascend 910）：专用的AI加速器，是计算集群的基础。
- **MindSpore框架**：一个开源的深度学习平台，支持**自动并行技术**，该技术结合了五种并行模式（数据并行、模型并行、流水线并行、优化器并行等），可将计算任务高效地分配到数千个节点上<sup>[\[3\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-tech_report_alpha-3)</sup>。

## 专用模型及其应用

### PanGu-Weather - 盘古气象

该系列最著名的模型之一是**盘古气象**（PanGu-Weather），这是一个基于深度学习的全球气象预测模型。2023年7月，关于该模型的论文发表在权威科学期刊《自然》（*Nature*）上<sup>[\[7\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-nature_pangu_news-7)</sup>。

该模型在预测精度上超越了**欧洲中期天气预报中心（ECMWF）**的传统数值方法，且速度要快得多。生成一个24小时的全球天气预报，该模型仅需**几秒钟**，而传统方法需要数小时的超级计算机计算，这意味着速度提升了约**10000倍**<sup>[\[7\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-nature_pangu_news-7)</sup>。2023年8月，盘古气象的预测结果被整合到 ECMWF 的服务中，用于实际的气象服务<sup>[\[8\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-nature_paper-8)</sup>。

### 行业应用

盘古模型已在30个行业的500多个场景中得到应用。部分示例如下：

- **农业**：中国农业科学院（CAAS）利用盘古开发了育种模型，成功培育出一种抗倒伏能力更强的实验性水稻品种<sup>[\[5\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-pangu5_techmag-5)</sup>。
- **石油和天然气行业**：中国石油天然气集团公司（CNPC）使用盘古模型自动检测管道的亚毫米级缺陷，将效率提升了约40%<sup>[\[9\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-cnpc_case-9)</sup>。
- **政务服务**：深圳市推出了智能助手**“小福”**，它基于超过20万份本地文件，为市民提供政务服务信息<sup>[\[4\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-pangu3_launch-4)</sup>。
- **药物研发**：**盘古药物分子**（PanGu Drug Molecule）模型被用于加速候选药物的筛选过程。据称，借助该模型发现了一类全新的抗生素，这是该领域40年来的首次突破<sup>[\[4\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-pangu3_launch-4)</sup>。

## 开源

2025年6月，华为宣布将盘古模型系列的部分模型开源（open-source）。开源的模型包括<sup>[\[10\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-opensource_2025-10)</sup>：

- **盘古密集模型7B**（PanGu Dense Model 7B，70亿参数）。
- **盘古Pro MoE模型72B**（PanGu Pro MoE Model 72B，720亿参数）。

此举旨在激励创新，并围绕**华为昇腾**（Huawei Ascend）硬件平台构建一个开放的生态系统，这是对全球AI领域竞争的战略回应<sup>[\[10\]](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_note-opensource_2025-10)</sup>。

## 参考文献

- Zeng, W.; et al. (2021). *PanGu‑α: Large‑Scale Autoregressive Pretrained Chinese Language Models*. <a href="https://s10251.pcdn.co/pdf/2021-zeng-pangu-alpha.pdf" class="external text" rel="nofollow">PDF</a>.
- Huawei (2021). *HDC.Cloud 2021: Huawei Releases Six Ground‑breaking Products to Supercharge the Cloud and Intelligent Transformation of Business*. <a href="https://www.huawei.com/en/news/2021/4/hdc-cloud-2021-six-products-cloud-intelligent-transformation" class="external text" rel="nofollow">Online news</a>.
- Huawei Cloud (2023). *Reshaping Industries with AI: Huawei Cloud Launches PanGu Models 3.0 and Ascend AI Cloud Services*. <a href="https://www.huaweicloud.com/intl/en-us/news/20230707180809498.html" class="external text" rel="nofollow">Online news</a>.
- Bi, K.; et al. (2023). *Accurate Medium‑Range Global Weather Forecasting with 3D Neural Networks*. *Nature*, 620, 560–566. <a href="https://www.nature.com/articles/s41586-023-06185-3" class="external text" rel="nofollow">DOI:10.1038/s41586‑023‑06185‑3</a>.
- Technology Magazine (2025). *What Huawei PanGu 5.5 Models Mean for Industrial AI*. <a href="https://technologymagazine.com/news/how-huawei-pangu-5-5-ai-models-transform-industry-operations" class="external text" rel="nofollow">Online article</a>.
- MindSpore Team (2021). *MindSpore: An All‑Scenario Deep Learning Computing Framework (White Paper v1.1)*. <a href="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/white_paper/MindSpore_white_paper_enV1.1.pdf" class="external text" rel="nofollow">PDF</a>.
- Zhang, S.; et al. (2024). *Ascend 910 NPU SoC Architecture for Large‑Scale AI Training*. *arXiv:2407.11888*. <a href="https://arxiv.org/html/2407.11888v1" class="external text" rel="nofollow">Online preprint</a>.
- AIbase News (2025). *Huawei Open Sources Dense PanGu 7B and Mixture‑of‑Experts PanGuPro 72B*. <a href="https://www.aibase.com/news/19320" class="external text" rel="nofollow">Online news</a>.
- CNPC & Huawei Cloud (2024). *Kunlun: Large‑Scale AI Model for Oil and Gas Pipeline Defect Detection*. <a href="https://www.cnpc.com.cn/en/RandDProgressin2024/202502/4bf8723a4b194ebeb2b6212579ab4ffa.shtml" class="external text" rel="nofollow">Online case study</a>.
- MindSpore Docs (2024). *Automatic Parallel — Five‑Mode Hybrid Strategy in MindSpore*. <a href="https://www.mindspore.cn/tutorials/experts/en/r2.3.0rc1/parallel/auto_parallel.html" class="external text" rel="nofollow">Online documentation</a>.
- Press, O.; et al. (2021). *Train Short, Test Long: Attention with Linear Biases Enables Input‑Length Extrapolation*. <a href="https://arxiv.org/abs/2108.12409" class="external text" rel="nofollow">arXiv:2108.12409</a>.
- Law, M. (2025). *How Huawei PanGu 5.5 AI Models Transform Industry Operations*. *AI Magazine*. <a href="https://aimagazine.com/articles/how-huawei-pangu-5-5-ai-models-transform-industry-operations" class="external text" rel="nofollow">Online article</a>.

## 注释

1.  <span id="cite_note-hdc2021-1">↑ <sup>[1.0](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-hdc2021_1-0)</sup> <sup>[1.1](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-hdc2021_1-1)</sup> “HDC.Cloud 2021: Huawei Releases Six Groundbreaking Products to Supercharge the Cloud and Intelligent Transformation of Business”. *Huawei*. <a href="https://www.huawei.com/en/news/2021/4/hdc-cloud-2021-six-products-cloud-intelligent-transformation" class="external autonumber" rel="nofollow">[1]</a></span>
2.  <span id="cite_note-aibusiness_2021-2">[↑](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-aibusiness_2021_2-0) Wodecki, Ben (27 Apr 2021). «Huawei has created the world's largest Chinese language model». *AI Business*. <a href="https://aibusiness.com/nlp/huawei-has-created-the-world-s-largest-chinese-language-model" class="external autonumber" rel="nofollow">[2]</a></span>
3.  <span id="cite_note-tech_report_alpha-3">↑ <sup>[3.0](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-tech_report_alpha_3-0)</sup> <sup>[3.1](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-tech_report_alpha_3-1)</sup> <sup>[3.2](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-tech_report_alpha_3-2)</sup> Zeng, Wei, et al. (Apr 2021). «PanGu-α: Large-scale Autoregressive Pretrained Chinese Language Models». *Technical Report*. <a href="https://s10251.pcdn.co/pdf/2021-zeng-pangu-alpha.pdf" class="external autonumber" rel="nofollow">[3]</a></span>
4.  <span id="cite_note-pangu3_launch-4">↑ <sup>[4.0](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-pangu3_launch_4-0)</sup> <sup>[4.1](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-pangu3_launch_4-1)</sup> <sup>[4.2](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-pangu3_launch_4-2)</sup> «Reshaping Industries with AI: Huawei Cloud Launches Pangu Models 3.0 and Ascend AI Cloud Services». *HUAWEI CLOUD*. 7 Jul 2023. <a href="https://www.huaweicloud.com/intl/en-us/news/20230707180809498.html" class="external autonumber" rel="nofollow">[4]</a></span>
5.  <span id="cite_note-pangu5_techmag-5">↑ <sup>[5.0](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-pangu5_techmag_5-0)</sup> <sup>[5.1](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-pangu5_techmag_5-1)</sup> Law, Marcus (23 Jun 2025). «What Huawei Pangu 5.5 Models Mean for Industrial AI». *Technology Magazine*. <a href="https://technologymagazine.com/news/how-huawei-pangu-5-5-ai-models-transform-industry-operations" class="external autonumber" rel="nofollow">[5]</a></span>
6.  <span id="cite_note-aimagazine_5_5-6">[↑](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-aimagazine_5_5_6-0) «How Huawei Pangu 5.5 AI Models Transform Industry Operations». *AI Magazine*. <a href="https://aimagazine.com/articles/how-huawei-pangu-5-5-ai-models-transform-industry-operations" class="external autonumber" rel="nofollow">[6]</a></span>
7.  <span id="cite_note-nature_pangu_news-7">↑ <sup>[7.0](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-nature_pangu_news_7-0)</sup> <sup>[7.1](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-nature_pangu_news_7-1)</sup> «Prestigious science journal Nature publishes paper about Pangu Weather AI Model authored by HUAWEI CLOUD researchers». *Huawei News*. 6 Jul 2023. <a href="https://www.huawei.com/en/news/2023/7/pangu-ai-model-nature-publish" class="external autonumber" rel="nofollow">[7]</a></span>
8.  <span id="cite_note-nature_paper-8">[↑](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-nature_paper_8-0) Bi, Kaifeng, et al. (2023). «Accurate medium-range global weather forecasting with 3D neural networks». *Nature*. <a href="https://www.nature.com/articles/s41586-023-06185-3" class="external autonumber" rel="nofollow">[8]</a></span>
9.  <span id="cite_note-cnpc_case-9">[↑](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-cnpc_case_9-0) «CNPC and Huawei Cloud Jointly Launch the "Kunlun" Model for the Oil and Gas Industry».</span>
10. <span id="cite_note-opensource_2025-10">↑ <sup>[10.0](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-opensource_2025_10-0)</sup> <sup>[10.1](https://systems-analysis.info/int/PanGu_(Huawei)_%E2%80%94_%E5%8D%8E%E4%B8%BA%E7%9B%98%E5%8F%A4#cite_ref-opensource_2025_10-1)</sup> «Huawei Open Sources Dense Pangu 7B and Mixture of Experts Model with 72B Parameters». *Albase News*. 30 Jun 2025. <a href="https://www.aibase.com/news/19320" class="external autonumber" rel="nofollow">[9]</a></span>
