Pub-AI logoPub-AI: AI in Science
Domain DiscoveryOpen Source🔓 open source🔗 source linked

Pangu-Weather

Huawei's 3D transformer-based weather forecasting model — one of the first AI models to outperform ECMWF HRES, with extremely fast inference and a 3D atmospheric architecture that sets it apart from graph-based competitors.

Last verified: July 2026

What it does

Pangu-Weather is an AI global weather forecasting model from Huawei Cloud, published in Nature (2023). It uses a 3D Earth Specific Transformer (3DEST) architecture that explicitly represents the atmosphere in three spatial dimensions and multiple pressure levels simultaneously — unlike graph neural network approaches (GraphCast) that treat pressure levels more independently.

Pangu-Weather generates deterministic 10-day global forecasts from ERA5-initialized inputs. Its key distinguishing features are extremely fast inference (milliseconds to seconds per forecast step on appropriate hardware) and strong performance at short-to-medium lead times (24–72 hours) relative to ECMWF HRES.

Four separate model weights are provided for different forecast horizons (1h, 3h, 6h, 24h), allowing efficient multi-step forecasting by using the appropriate model at each step rather than chaining a single model repeatedly.

Best for

Research requiring fast, high-cadence deterministic forecasts: operational workflows that need to generate forecasts rapidly, comparison studies of AI weather model architectures, and short-to-medium range research where Pangu-Weather’s strong 24–72 hour performance is most relevant. Also suited to research groups exploring transformer architectures for geophysical applications.

Pricing

Free and open-source. Code and pretrained model weights released on GitHub. Requires GPU infrastructure to run; no hosted inference API from Huawei.

Strengths

  • Fastest inference of the major open-source AI weather models — significantly faster than GraphCast, which uses a graph neural network requiring more computation per step
  • 3D atmospheric architecture: the 3D transformer treats vertical pressure levels and horizontal grid jointly, which may capture vertical coupling in the atmosphere more directly than approaches that treat levels separately
  • Multi-horizon model weights: using the 24h model for 24-hour steps rather than iterating the 6h model four times avoids compounding of error across steps
  • Published in Nature with reproducible benchmark comparisons against ECMWF HRES
  • Strong short-range performance (1–3 day forecasts), where small errors matter most for applications

Limitations

  • Deterministic only — produces a single forecast with no uncertainty information; use GenCast when ensemble output is needed
  • 7-day practical skill limit — at longer lead times (beyond 7–10 days), accuracy degrades more rapidly than GraphCast and GenCast, which are optimized for medium-range forecasting
  • Requires ERA5-formatted input data, which involves preprocessing steps not needed for web-based forecast access
  • Self-published performance comparisons beyond the original Nature paper (particularly vs. newer AI models) are from Huawei and have not been independently benchmarked — treat these with appropriate skepticism
  • No ocean, land, or sea-ice coupling — atmosphere-only, like the other AI weather models

How it compares

vs. Key difference
GraphCast GraphCast uses a graph neural network; Pangu-Weather uses a 3D transformer. Both are deterministic. GraphCast is stronger at longer lead times; Pangu-Weather has faster inference
GenCast GenCast produces probabilistic ensemble output; Pangu-Weather is deterministic. GenCast is the choice when uncertainty quantification matters
ECMWF HRES HRES uses physics-based NWP; Pangu-Weather is data-driven. Pangu-Weather matches or exceeds HRES on standard metrics; HRES is better-understood for rare events

Related content

← Back to Tools Directory