Huawei PanGu
Huawei's multimodal large model family for industry, from 1B to trillion-scale
Huawei PanGu (openPangu) is a multimodal large language model family from Huawei, officially launched in July 2023 and named after the Chinese mythological figure Pangu. It spans embedded to trillion-parameter versions, targets industry tasks such as finance, mining and weather, and was open-sourced in 2025.
- PanGu-Σ has 1.085 trillion parameters (April 2023 paper)
- Trained on 512 Ascend 910 chips for over 100 days
- Covers 329 billion tokens across 40+ languages
- openPangu 7B and 72B Pro MoE weights released June 30, 2025
- PanGu-Weather model used by ECMWF for global forecasting
What it is
Huawei PanGu, also called openPangu and 盘古大模型, is a multimodal large language model family built by Huawei. It was officially launched in July 2023 and takes its name from Pangu, the primordial figure of Chinese mythology said to have formed the world. The family ranges from a 1-billion-parameter embedded version to trillion-level editions. Rather than aiming at casual chat, it is pitched at enterprise and public-sector work: government, finance, manufacturing, mining and meteorology. Huawei Cloud offers it as a platform where customers adapt models to their own data and tasks.
How it came about
Huawei published work on PanGu-Σ in April 2023, describing a 1.085-trillion-parameter model trained on 512 Ascend 910 chips. PanGu 3.0 was unveiled at the Huawei Developer Conference on July 7, 2023, aimed at industry sectors. The following month Huawei launched the Celia virtual assistant, set to ship with HarmonyOS 4.0. PanGu 5.0 arrived at HDC 2024 on June 21 alongside HarmonyOS NEXT. On June 30, 2025 Huawei open-sourced models as openPangu, and at Huawei Connect 2025 on September 29 it set out a plan to open its full software stack by the end of that year.
Inside the model
PanGu-Σ uses a Transformer decoder with a technique called Random Routed Experts (RRE), part of the mixture-of-experts approach where only some sub-networks activate for a given input. This design lets sub-models be pulled out for uses such as conversation, translation and code, and gave 6.3 times the training throughput of comparable MoE setups. Training ran over 100 days on 512 Ascend 910 chips in Huawei's MindSpore framework, covering 329 billion tokens from 40 domains. PanGu 3.0 uses a 5+N+X layout: five base models (NLP, vision, multimodal, prediction, scientific computing), then industry models, then scenario-specific ones.
What it can do
PanGu-Σ was reported to beat earlier state-of-the-art results on 16 Chinese tasks in a zero-shot setting, and to handle few-shot understanding, open-domain discussion, question answering, translation and code generation across 40-plus languages. PanGu 5.0 added multi-style modelling and generative features for text, code and images. The PanGu-Weather model, built with MindSpore and Huawei Cloud, supplies global forecasting data through the ECMWF site.
How to use it
PanGu is offered through Huawei Cloud, with customers training on their own data. On June 30, 2025 Huawei released openPangu weights, including a 7-billion-parameter model and a 72-billion-parameter openPangu Pro MoE, plus inference technology tuned for Ascend chips. The licence is listed as open source, and code is hosted on GitCode. Huawei announced a roadmap to open its MindSpore toolchains, openPangu models and CANN interfaces by December 31, 2025.
Safety and testing
Huawei's response to the 2025 similarity claims stated that PanGu is a foundational model developed on Ascend hardware and not incrementally trained on other models, and that attributions were made in line with open-source licences. No separate safety evaluation or red-teaming details are given here.
Reception and debate
In July 2025 researchers posting on GitHub alleged a very high similarity between the attention parameter distribution of Pangu Pro MoE and Alibaba's Qwen, based on model fingerprinting. Huawei Noah's Ark Lab rejected the claim the next day, saying the model was self-developed and that compliant attributions were made. The repository carrying the accusation was later deleted.
Why it matters
PanGu pushed Huawei's Ascend chips and MindSpore framework into large-model work, with a weather model adopted by the European Centre for Medium-Range Weather Forecasts. Its 5+N+X tiering and the E, P, U and S series spread the family from phone-scale to trillion-parameter use, and the 2025 openPangu release put Huawei weights into the open community.
Questions people ask about Huawei PanGu
What is Huawei PanGu?
PanGu, also known as openPangu, is a multimodal large language model family developed by Huawei and officially launched in July 2023. It is aimed mainly at enterprise and public-sector tasks such as finance, manufacturing, mining and weather forecasting rather than general chat.
How large are the models?
The family spans several scales. PanGu-Σ, described in an April 2023 paper, has 1.085 trillion parameters. PanGu 5.0 splits into an E series at 1 billion, a P series at 10 billion, a U series at 135 and 230 billion, and an S series at trillion level.
Can I download the weights?
Yes. On June 30, 2025 Huawei released openPangu models, including a 7-billion-parameter version and a 72-billion-parameter openPangu Pro MoE, with inference technology optimised for Ascend AI accelerator chips. The licence is listed as open source.
What hardware does it run on?
PanGu-Σ was trained inside Huawei's MindSpore framework on a cluster of 512 Ascend 910 AI accelerator chips, taking more than 100 days. The 2025 openPangu release included inference technology tuned for those same Ascend chips.
Was there any controversy?
In July 2025 researchers on GitHub alleged high similarity between Pangu Pro MoE and Alibaba's Qwen using model fingerprinting. Huawei Noah's Ark Lab said PanGu was self-developed on Ascend hardware, not built on other models, and that licence attributions were compliant. The accusation repository was later deleted.
Good reading and sources
The facts on this page come from the Wikipedia article "Huawei PanGu", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.