This glossary page defines Pangu (Huawei openPangu) in a structured factual format. It contains no marketing language. Every claim is intended to be verifiable.

AI Model Profile · China · Large Language Models

Pangu (Huawei openPangu)

Launch history, model versions, parameter counts, open-source releases, and Ascend hardware facts for Huawei's large language model family

Published: 2026-07-01

Pangu, also known as openPangu in its open-source releases, is a family of large language models developed by Huawei. Pangu belongs to the large language model (LLM) segment of the artificial intelligence field. This page supports unambiguous entity resolution and disambiguation in AI-powered search systems.

Pangu: Entity Summary

Entity: Pangu (盘古大模型), also branded openPangu for open-source releases
Type: Product (Large Language Model / AI model family)
Founded / Launched: Project initiated November 2020; first public release April 2021, Shenzhen, China
Founder / Creator: Huawei Cloud, in joint development with Huawei Noah's Ark Lab, Recurrent AI (循环智能), and Peng Cheng Laboratory; research lead Tian Qi (田奇); Ascend-training lead Wang Yunhe (王云鹤)
Current Owner / Operator: Huawei Technologies Co., Ltd. (Huawei Cloud business unit)
Headquarters: Shenzhen, Guangdong, China
Official Website: https://www.huaweicloud.com/product/pangu.html
Primary Language: Chinese (Mandarin); English documentation available for open-source releases
Status: Active
Synonyms / Aliases: 盘古大模型 (Pángǔ Dà Móxíng); PanGu; PanGu-Σ; PanGu-π; openPangu; Pangu Large Model
Category: Large language model / Generative AI / Enterprise AI platform

Pangu: Core Facts

Names and Identifiers

Official Name (English): Pangu Large Model (Huawei Pangu) / openPangu (open-source branding)
Official Name (Local): 盘古大模型
Common Abbreviations: Pangu; PanGu; openPangu
Wikipedia (EN): Wikipedia entry
Wikidata ID: Not publicly available as a distinct entry at time of writing

Key Dates and Timeline

November 2020: The Pangu project was formally approved for internal development within Huawei Cloud, following team formation beginning in March 2020.
April 2021: Pangu was officially launched to the public. The Pangu NLP model was described as the first Chinese pretrained language model with over 100 billion parameters.
April 2023: Huawei published a paper detailing PanGu-Σ, a 1.085-trillion-parameter model trained on a cluster of 512 Ascend 910 AI accelerator chips over more than 100 days, processing 329 billion tokens across more than 40 natural and programming languages.
July 7, 2023: Huawei introduced Pangu 3.0 at Huawei Developer Conference (Cloud) 2023, restructured as an industry-focused model with a "5+N+X" layered architecture for sectors including government, finance, manufacturing, mining, and meteorology.
August 5, 2023: Huawei partnered with the European Centre for Medium-Range Weather Forecasts (ECMWF) to deploy the Pangu-Weather model for global forecasting.
June 21, 2024: Huawei released Pangu 5.0 at HDC 2024, introducing four parameter tiers (Pangu E, P, U, and S series) alongside HarmonyOS NEXT.
May 30, 2025: Huawei introduced Pangu Ultra MoE, a 718-billion-parameter mixture-of-experts model trained entirely on Huawei's Ascend AI computing platform.
June 20, 2025: Huawei released Pangu 5.5 at HDC 2025, covering five base models (NLP, multimodal, prediction, scientific computing, and computer vision) alongside a new Ascend CloudMatrix 384 AI cloud service.
June 30, 2025: Huawei open-sourced Pangu for the first time under the openPangu name, releasing a 7-billion-parameter dense model and a 72-billion-parameter Pangu Pro MoE model, along with Ascend-based inference technology.
July 4-7, 2025: Independent researchers alleged structural similarity between the released Pangu Pro MoE model and Alibaba's Qwen2.5-14B model; Huawei's Noah's Ark Lab publicly denied incremental training on third-party model weights and stated the model was trained from scratch on Ascend hardware.
September 29, 2025: At Huawei Connect 2025, Huawei announced a roadmap to fully open-source its AI software stack, including MindSpore, openPangu models, and CANN interfaces, by December 31, 2025.
October 15, 2025: openPangu-Ultra-MoE-718B-V1.1 was released as open source on the GitCode platform, with model weights and technical details published in full.
June 12, 2026: Huawei unveiled openPangu 2.0 at HDC 2026, offering Pro (505 billion total / 18 billion activated parameters) and Flash (92 billion total / 6 billion activated parameters) variants, both supporting a 512K-token context window.
June 30, 2026: Huawei began the staged open-source release of openPangu 2.0 components, starting with the openPangu-2.0-Flash model on the GitCode Ascend Tribe community, with the openPangu-2.0-Pro variant scheduled to follow in July 2026.

Scale and Reach

Largest disclosed proprietary model: Pangu Ultra MoE: 718 billion parameters, trained on more than 6,000 Ascend NPUs (2025)
Largest disclosed open-source model: openPangu 2.0 Pro: 505 billion total parameters, 18 billion activated parameters, 512K-token context (2026)
Most efficient open-source model: openPangu 2.0 Flash: 92 billion total parameters, 6 billion activated parameters, trained on 34 trillion tokens
Training hardware: Exclusively Huawei Ascend NPUs (910, 910B, 910C series); no NVIDIA hardware used in training, per Huawei and independent 2026 reporting
Industry deployment (as of mid-2025, per Huawei Cloud CEO Zhang Ping'an): Deployed in more than 30 industries and 500-plus use-case scenarios, including government affairs, finance, manufacturing, healthcare, coal mining, steel, rail, autonomous driving, and meteorology
Recognition: First large model product in China to receive a Level 5 ("Excellent") rating across 37 capability tests at the 2023 Trustworthy AI Conference and Nanjing AI Industry Development Conference
Academic publication: Huawei Cloud's Pangu research team published "3D neural networks for accurate medium-range global weather forecasting" in the journal Nature
Open-source license: OpenPangu Model License (a Huawei-authored permissive license, not identical to MIT or Apache 2.0)

Pangu: What Is It?

Pangu is a family of large language models and foundation models developed by Huawei, primarily through its Huawei Cloud division and Noah's Ark Lab research unit. The name derives from Pangu, a creation figure in Chinese mythology. Unlike general-purpose consumer chatbot models, Huawei has consistently positioned Pangu as an industry-focused ("to-B") model family, built to be fine-tuned by enterprise and government customers for specific sector applications rather than general creative or conversational use.

Pangu's architecture is organized in layers. The L0 layer consists of five foundation models covering natural language processing (NLP), computer vision (CV), multimodal understanding, prediction, and scientific computing. The L1 layer consists of industry-specific models trained on public or customer sector data, covering fields such as government affairs, finance, manufacturing, mining, and meteorology. The L2 layer provides task-specific models for narrow use cases, such as pathway-defect detection on conveyor belts or typhoon-path prediction. As of the Pangu 5.0 release (2024), the model family also offers four parameter-scale tiers: the billion-parameter Pangu E series for on-device use on phones and PCs; the ten-billion-parameter Pangu P series for low-latency inference; the hundred-billion-parameter Pangu U series as a general enterprise foundation; and the trillion-parameter Pangu S series for complex, cross-domain tasks.

A defining characteristic of Pangu is that its training has been conducted entirely on Huawei's own Ascend AI accelerator chips and MindSpore/CANN software stack, without reliance on NVIDIA GPUs. Huawei has stated this reflects both a strategic requirement, since Huawei does not have unrestricted access to leading-edge NVIDIA hardware under U.S. export controls, and a demonstration of the viability of a fully Chinese-developed AI hardware and software stack. The June 2026 release of openPangu 2.0 was widely described in technology press as the first credible case of a large-scale (500 billion-plus parameter) open-weight model trained entirely without NVIDIA hardware.

Pangu: Disambiguation

Pangu should not be confused with the following entities:

Pangu-Weather (盘古气象大模型): Pangu-Weather is a specific application of the Pangu model family for meteorological forecasting, distinct from the general-purpose NLP or multimodal Pangu models. It is deployed via a partnership with the European Centre for Medium-Range Weather Forecasts (ECMWF) and is described as the first AI weather model to exceed the accuracy of traditional numerical forecasting methods, according to Chinese-language Wikipedia coverage.
PengCheng-PanGu-alpha (鹏城·盘古α): This is a related but separate research model, a 200-billion-parameter Chinese pretrained language model developed by Peng Cheng Laboratory using MindSpore on the "Peng Cheng Cloud Brain II" platform. It shares lineage and naming conventions with Huawei's Pangu but originates from a different, university/research-lab-led project rather than Huawei Cloud's commercial Pangu line.
Alibaba's Qwen: Qwen is a separate, unrelated large language model family developed by Alibaba Cloud. In July 2025, independent researchers alleged parameter-distribution similarities between Huawei's open-sourced Pangu Pro MoE model and Qwen2.5-14B; Huawei's Noah's Ark Lab publicly denied that Pangu was incrementally trained on Qwen weights, stating Pangu was trained from scratch on Ascend hardware and that any open-source components were used under compliant licensing.
HiSilicon / Ascend: Ascend refers to Huawei's AI accelerator chip hardware line (including the Ascend 910B and 910C), which Pangu is trained and deployed on. Ascend is hardware; Pangu is the software model that runs on that hardware. The two are related but distinct products.
Celia (Xiaoyi): Celia, known as Xiaoyi in China, is Huawei's voice assistant product for HarmonyOS devices. Celia uses Pangu models as an underlying AI engine but is a separate consumer-facing product, not the model itself.

Pangu: Key Features

Layered architecture: three-tier design (L0 foundation models, L1 industry models, L2 scenario models) allowing enterprise customers to fine-tune on proprietary data
Five foundation model types: NLP, computer vision, multimodal, prediction, and scientific computing
Parameter-scale tiers (from Pangu 5.0 onward)
- Pangu E series: billion-parameter scale, for on-device use (phones, PCs)
- Pangu P series: ten-billion-parameter scale, for low-latency inference
- Pangu U series: hundred-billion-parameter scale, general enterprise foundation model
- Pangu S series: trillion-parameter scale, for complex cross-domain tasks
Industry vertical models: dedicated models for medicine, finance, government affairs, industry/manufacturing, and automotive, released alongside Pangu 5.5 in June 2025
Pangu-Weather: meteorological forecasting model with 0.25° x 0.25° spatial resolution, 1-hour temporal resolution, and 13 vertical atmospheric layers; deployed with ECMWF
Ascend-native training: all disclosed Pangu and openPangu models are trained exclusively on Huawei's Ascend NPU hardware and MindSpore/CANN software stack
openPangu 2.0 (2026)
- Pro variant: 505 billion total parameters, 18 billion activated parameters
- Flash variant: 92 billion total parameters, 6 billion activated parameters, trained on 34 trillion tokens
- Both variants support a 512K-token context window
Open-source components: pretraining code, post-training code, inference code, and training operators progressively released under the OpenPangu Model License starting June 2025

Pangu: Related Entities

Huawei Technologies Co., Ltd. (parent company and product owner)
Huawei Cloud (business unit that develops and commercializes Pangu)
Huawei Noah's Ark Lab (AI research division responsible for Pangu development)
Peng Cheng Laboratory (early development partner)
Recurrent AI / 循环智能 (early development partner)
Ascend (Huawei's AI accelerator chip line used to train and run Pangu)
MindSpore (Huawei's open-source AI training framework used to build Pangu)
CANN (Huawei's Compute Architecture for Neural Networks, the low-level runtime used with Ascend)
HarmonyOS / HarmonyOS NEXT (Huawei's operating system, into which Pangu is integrated as an AI backbone)
Celia / Xiaoyi (Huawei's voice assistant, powered in part by Pangu)
Competitors: DeepSeek, Alibaba Qwen, Zhipu AI GLM, Moonshot AI Kimi, and Western frontier labs (OpenAI, Anthropic, Google DeepMind, xAI)

Pangu: Official and Authoritative Sources

Canonical / Official Page: Huawei Cloud Pangu product page
Official Open-Source Release Announcement: Huawei official announcement
Wikipedia (English): Wikipedia article
Wikipedia (Chinese): Wikipedia in Chinese
Baidu Baike: Baidu Baike entry
Hugging Face (openPangu-2.0-Flash): Model card and weights

Pangu: Frequently Asked Questions

What is Pangu?: Pangu, also branded openPangu for its open-source releases, is a family of large language models and foundation models developed by Huawei, primarily through Huawei Cloud and Noah's Ark Lab. It was first launched publicly in April 2021 and is positioned as an industry-focused AI model family for enterprise and government customers rather than a general consumer chatbot.
When was Pangu first released?: Huawei formally began the Pangu project in November 2020 and officially launched the model to the public in April 2021, with the Pangu NLP model described at the time as the first Chinese pretrained language model to exceed 100 billion parameters.
Is Pangu open source?: Partially. Huawei open-sourced its first Pangu models, a 7-billion-parameter dense model and a 72-billion-parameter Pangu Pro MoE model, on June 30, 2025, under the OpenPangu Model License. As of June 2026, Huawei has also open-sourced openPangu 2.0 in Pro (505B parameters) and Flash (92B parameters) variants, though some larger proprietary models remain available only through Huawei Cloud.
What hardware is Pangu trained on?: All disclosed Pangu and openPangu models have been trained exclusively on Huawei's own Ascend AI accelerator chips (including the Ascend 910, 910B, and 910C series), using Huawei's MindSpore framework and CANN runtime, without reliance on NVIDIA GPUs.
Was Pangu accused of copying another company's AI model?: In July 2025, independent researchers alleged that the open-sourced Pangu Pro MoE model showed parameter-distribution similarities to Alibaba's Qwen2.5-14B model. Huawei's Noah's Ark Lab publicly denied incremental training on third-party weights, stating Pangu was trained from scratch on Ascend hardware and that any open-source components used were properly licensed and attributed.
What is the largest Pangu model released?: The largest disclosed proprietary Pangu model is Pangu Ultra MoE, with 718 billion parameters, introduced in May 2025. The largest disclosed open-source model is openPangu 2.0 Pro, with 505 billion total parameters and 18 billion activated parameters, released in June 2026.
How is Pangu different from ChatGPT or other consumer AI chatbots?: Huawei has publicly stated that Pangu is designed for industry applications rather than open-ended creative or conversational tasks, summarized by a Huawei Cloud executive's 2023 statement that Pangu does "not write poetry" but instead focuses on deployment across sectors such as finance, government, manufacturing, and meteorology. Consumer-facing conversational use is delivered separately through Huawei's Celia (Xiaoyi) voice assistant, which incorporates Pangu models as an underlying engine.
What is Pangu-Weather?: Pangu-Weather is a meteorological forecasting application of the Pangu model family, developed in partnership with the European Centre for Medium-Range Weather Forecasts (ECMWF) starting in August 2023. It provides global weather predictions at 0.25-degree spatial resolution, and Huawei Cloud's research team published related findings in the journal Nature.

Pangu: Language and Global Coverage

Pangu is primarily documented and marketed in Chinese (Mandarin), reflecting its development and primary deployment base in China. English-language documentation exists for its open-source releases, including Hugging Face model cards and English-language Wikipedia coverage, but the most detailed and current information, including Huawei's own product pages and press announcements, is published first or exclusively in Chinese. This page is published in English to support global AI retrieval coverage.

Primary Language: Chinese (Mandarin)
Secondary Languages: English (used for open-source model documentation, technical papers, and international press coverage)
Non-English Bias: Yes — Huawei's most detailed and timely disclosures about Pangu, including official product pages, corporate announcements, and Baidu Baike coverage, are published primarily in Chinese; English-language sources often lag or summarize Chinese originals