# Pangu-Weather by Huawei

Canonical page: https://technooptimist.io/ai/pangu-weather
Frontier: [AI](https://technooptimist.io/ai.md) · Field: Weather models · Company: [Huawei](https://technooptimist.io/companies/huawei.md)
Stage: Deploy (4 of 5: Research → Proto → Pilot → Deploy → Scaling)
Updated: last update 2024-03-23, facts checked 2026-10-11
Goal: Replace hours of supercomputer physics with a learned model that forecasts global weather in seconds, at least as accurately as the best traditional forecast, so far more forecasts can be run and shared.
Status: Published in Nature in 2023; charted by ECMWF and behind Shenzhen's 3 km Zhiji regional forecasts.

## Milestones

| Status | Milestone | Date | Slip |
| --- | --- | --- | --- |
| Done | Preprint: first AI model to beat physics forecasts on all variables | 3 Nov 2022 |   |
| Done | Published in Nature | 5 Jul 2023 |   |
| Done | Forecasts public on the ECMWF website | Aug 2023 |   |
| Done | Zhiji 3 km regional model in use in Shenzhen | 23 Mar 2024 |   |

## Most important updates

- **23 Mar 2024**: Shenzhen Meteorological Bureau and Huawei Cloud launch Zhiji, a 3 km regional AI forecast built on Pangu-Weather. Shenzhen's weather bureau and Huawei Cloud put Zhiji into use, a regional model built on Pangu-Weather and fine-tuned on local data. ([our story](https://technooptimist.io/stories/shenzhen-meteorological-bureau-and-huawei-cloud-launch-zhiji-a-3-km-164adcc0.md) · [original source](https://www.huaweicloud.com/intl/en-us/news/20240325084358116.html))
- **5 Jul 2023**: Huawei Cloud's Pangu-Weather AI model, faster and more accurate than physics forecasts in tests, published in Nature. Nature published Huawei Cloud's Pangu-Weather, an AI model trained on 39 years of global weather data that beat ECMWF's operational physics forecast on every variable tested, for forecasts from 1 hour to 7 days. ([our story](https://technooptimist.io/stories/huawei-cloud-s-pangu-weather-ai-model-faster-and-more-accurate-than-a2f97566.md) · [original source](https://www.nature.com/articles/s41586-023-06185-3))

## Current obstacles

- **Rain is the weak spot**: After Zhiji's 2024 trial, Shenzhen's team named precipitation forecasts as the part that still needed improving.
- **Needs a physics starting point**: On ECMWF's charts it starts from ECMWF's physics-based analysis of today's weather, so it still leans on the system it aims to replace.

## Physics limits

- **Learned only from the past**: It knows only the weather of 1979–2021, so extremes beyond that record are where it is least tested; chaos in the atmosphere caps useful forecasts at about two weeks.

## How it works

### 3D grid: Height as a third dimension

Its 3D Earth-Specific Transformer treats the stack of pressure levels as a cube, so it learns how air at different heights moves together.

### Time steps: Big jumps, fewer errors

Several networks each forecast a different jump ahead; taking the biggest jumps first means fewer steps, so small errors pile up less.

### Training: Learned from four decades of weather

Trained on ERA5, ECMWF's hour-by-hour reconstruction of global weather since 1979, instead of solving physics equations.

### Regional: Zhiji: Shenzhen at 3 km

Fine-tuned on regional data, Zhiji gives 5-day temperature, rain and wind forecasts for Shenzhen at 3 km, versus ~25 km globally.

## Spec sheet

| Spec | Pangu-Weather | Provenance |
| --- | --- | --- |
| Global grid | 0.25° (~25 km) | reported |
| Forecast range tested | 1 hour to 7 days | reported |
| Training data | Hourly ERA5, 1979–2021 | reported |
| Model size (all networks) | ~256M parameters | reported |
| 24-hour global forecast | ~1.4 s on one V100 GPU | reported |
| Speed vs. ECMWF IFS | >10,000x faster | reported |
| Zhiji regional model | 3 km, 5-day forecasts | reported |

Provenance: reported = stated by the company; estimated = our estimate; sample = a sample figure.

## About Huawei

Huawei designs semiconductors, cloud systems and telecoms gear in 170 countries. Founded 1987, it grew to lead telecoms infrastructure. Now competing directly with NVIDIA in AI accelerators with the Ascend chip series.

- Operates in 170+ countries across cloud, AI, telecoms, consumer devices and transport systems.
- Ascend roadmap: 950/960/970 series at one-year cycles, each doubling compute. 950PR Q1 2026, 950DT Q4 with proprietary HBM.
- Atlas 950 SuperPoD (8,192 NPUs) delivers triple H20 inference performance at quarter the cost, shown at MWC 2026.

- Founded: 1987
- Headquarters: CN
- Status: private
- Staff: ~213,000
- Website: https://www.huawei.com/en/
- People: Ren Zhengfei (Founder and CEO)
- Partners, customers, investors: DeepSeek (customer); Baidu (customer); Alibaba (customer); SMIC (supplier)

Company page: https://technooptimist.io/companies/huawei.md

## Related programs

More from Huawei:

- [Ascend 950 / Atlas 950 SuperPoD](https://technooptimist.io/computing/ascend-950-atlas-950-superpod.md) (Proto)
- [Qiankun ADS (Autonomous Driving Solution)](https://technooptimist.io/vehicles/qiankun-ads.md) (Deploy)
- [China EUV Lithography](https://technooptimist.io/computing/china-euv-lithography.md) (Proto)
- [HIMA (Harmony Intelligent Mobility Alliance)](https://technooptimist.io/vehicles/hima.md) (Scaling)

Everything in AI: https://technooptimist.io/ai.md

---

The Techno Optimist tracks every frontier of technology, program by program: https://technooptimist.io/ (a Markdown version of any page: add .md to its URL; site map for agents: https://technooptimist.io/llms.txt).
Data API (JSON, realtime stream): https://technooptimist.io/developers. People read 5 pages free, then Reader $10/month; these Markdown pages are open to agents.