#2323rd of 42 in Open modelsFor science, health and research
weathernext
JAX code and weights for DeepMind's WeatherNext 2 and cyclone forecast models
Live demo ↗
(opens in a new tab)Documentation ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 7.7k
- License
- Apache-2.0
- Last commit
- Sep 2026
- Last release
- Aug 2026
- Language
- Python
6,509 on 1 Apr 20267,719 on 21 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Research code to run WeatherNext 2, WeatherNext Cyclones, and the earlier GraphCast and GenCast models for global medium-range weather and tropical cyclone forecasting. Pretrained weights are in a Google Cloud bucket, and a Colab notebook covers loading weights, autoregressive rollouts, plotting and the cyclone tracker. Full-size 0.25° models need an H100 or TPU; the 1° Mini models are meant for smaller hardware.
Who it is for: Weather and climate ML researchers running or fine-tuning global forecast models
Strengths
- Includes weights for operational 0.25° models and checkpoints that reproduce the Nature cyclone paper
- Colab notebook defaults to the 1° Mini model on a free v5e-1 TPU runtime
- Also contains GraphCast and GenCast code plus shared rollout and loss utilities
- Includes training loss and gradient-step code, not just inference
Weaknesses
- README calls it research code with no API stability guarantees; pinning a release is advised
- Non-Mini models need an H100 or TPU v5p; Mini models run on a P100
- Optimised for TPU; GPU use requires switching the attention implementation
- Non-code materials are CC BY 4.0; only code and notebooks are Apache 2.0
What it needs
- GPU required
- Needs JAX, Google Cloud Storage (weights), ERA5 (training), ECMWF HRES (initial conditions), WeatherBench2
- Models: WeatherNext 2, WeatherNext Cyclones, WeatherNext Cyclones Mini, GraphCast, GenCast
Also in Open models
See all 42| Rank | Project | Score |
|---|---|---|
| 1 | llmfitRust tool that checks which LLMs fit your RAM, CPU and GPU | 87 out of 100 |
| 2 | layaSingle-pass multilingual classifier for typed choice, score and yes/no decisions | 86 out of 100 |
| 3 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 76 out of 100 |
| #4 | speechPyTorch toolkit for training and deploying ASR, TTS and speech LLM models | 73 out of 100 |
| #5 | yueLyrics-to-song model that writes an editable melody and chord score first | 69 out of 100 |