Pangu-Weather

license

CC-BY-NC-SA-4.0(non-commercial use only)

weights

open(four ungated ONNX checkpoints on Google Drive and Baidu Netdisk, roughly 1.1GB each)

data

documented-not-released(ERA5 over 39 years, about 60TB, with the lite variant's 2007-2017 subset spelled out

code

partial(CPU, GPU and iterative inference scripts plus algorithm pseudocode

The repository declares no license file and GitHub's license endpoint returns 404, so it would be wrong to read the code as permissive. The only grant is the README's, and it puts the trained parameters under CC BY-NC-SA 4.0 and forbids commercial use, which caps the score at restricted whatever else was published.

• https://api.github.com/repos/198808xc/Pangu-Weather/license recorded 2026-10-06

  HTTP 404, "Not Found" - the repository carries no LICENSE file, so no repository-level grant exists to fall back on.

• https://raw.githubusercontent.com/198808xc/Pangu-Weather/master/README.md recorded 2026-10-06

  "The trained parameters of Pangu-Weather were made available under the terms of the BY-NC-SA 4.0 license" and "The commercial use of these models is forbidden"; links four downloadable ONNX checkpoints; states training needs about 60TB of ERA5 and 3000 V100 GPU-days per model; and ships inference_cpu.py, inference_gpu.py, inference_iterative.py and pseudocode.py.

Weights are hosted on Google Drive and Baidu Netdisk with no download count, so GitHub stars are the fallback reach signal instead.

• https://api.github.com/repos/198808xc/Pangu-Weather recorded 2026-09-10

  stargazers_count: 1395; archived: false; fork: false; pushed_at: 2024-01-12.

One attribute, the temporal-scale span. Pseudocode is not an adaptation pathway, ERA5 and ECMWF HRES initial fields are one data system, and a deterministic forecast is a fixed target rather than a generated output.

• https://raw.githubusercontent.com/198808xc/Pangu-Weather/master/README.md recorded 2026-09-10

  Publishes 1-hour, 3-hour, 6-hour and 24-hour models and an iterative script that combines them into a week of forecasts; supports ERA5 and ECMWF initial fields only; and provides pseudocode rather than training code.