WeatherNext 2: Google DeepMind's AI Weather Model
/weathernext
WeatherNext 2 is Google DeepMind's global, medium-range atmospheric and cyclone forecasting model, released with weights, inference code, and a Colab demo under Apache 2.0 and CC BY 4.0. Reach for it to run or study an operational-grade AI weather model, or to pull its forecasts from Google Cloud, WeatherLab, or OpenMeteo. Skip it if you need a stable, supported API — the README calls this research code with no promised stability.
Understanding WeatherNext 2 and its Predecessors
WeatherNext 2 is the current generation of Google DeepMind's WeatherNext model family, and this repository is its primary home. It also hosts the two prior generations: WeatherNext Graph, published as GraphCast, a deterministic medium-range forecaster built on graph neural networks, and WeatherNext Gen, published as GenCast, a diffusion-based ensemble forecaster. All three share code and utilities in one codebase.
Key Capabilities of WeatherNext 2 Models
- ✓The WeatherNext2_<2025 checkpoint runs at 0.25° resolution (~30km), is fine-tuned on ECMWF HRES data, initializes directly from operational HRES conditions rather than ERA5 reanalysis, is trained on data through 2024, and is used operationally.
- ✓WN2 also forecasts 100m wind — the one functional difference from WeatherNext Cyclones. Both use the identical cyclone-forecasting algorithm but carry different weights from independent training runs.
- ✓Three WeatherNext Cyclones checkpoints (<2025, <2024, <2023) reproduce the paper's cyclone results at different training cutoffs; the <2025 checkpoint ran live during the 2025 Atlantic hurricane season under the public name FNV3 (NHC's postprocessed version is called GDMI).
- ✓Two WeatherNext Cyclones Mini checkpoints run at 1° resolution and trade accuracy for lower compute — the README says they aren't expected to match the larger models but do run on a single TPU or GPU.
- ✓A shared utils/ directory provides autoregressive rollout, input normalization, graph-building blocks, loss computation, and JAX-compatible xarray utilities used across every model generation in the repo.
- ✓The demo notebook includes a direct tracker that pulls cyclone track data straight out of model output.
Accessing WeatherNext 2 Forecast Data and Models
- •Pull daily WN2 forecast output without running any code, through Google Cloud (Earth Engine, BigQuery, Vertex AI), WeatherLab (cyclone tracks included), or OpenMeteo's API and interactive builder.
- •Run the Colab demo notebook on the free v5e-1 TPU runtime to generate a forecast from the default WeatherNext Cyclones Mini checkpoint.
- •Load different pretrained checkpoints — WN2, WeatherNext Cyclones, or the Mini variants — to compare the operational model against a fixed research-cutoff version.
- •Fine-tune or continue training with ERA5 and HRES data pulled through WeatherBench2 for research work.
- •Run the notebook's direct tracker on model output to extract cyclone track data for further analysis.
Setting Up WeatherNext 2 for Local Use
Install with pip install git+https://github.com/google-deepmind/[email protected], the README's own example, pinned to a release since this is unversioned research code with no API-stability guarantee. You then need to separately download the pretrained weights from the project's Google Cloud Bucket — they aren't bundled with the pip package. Full training additionally requires the ERA5 dataset from ECMWF, best pulled as Zarr through WeatherBench2, plus WeatherBench2's HRES data for operational fine-tuning.
Running WeatherNext 2 with the Colab Notebook
The README's recommended path is the interactive Colab notebook at docs/weathernext2/wn2_demo.ipynb, which defaults to WeatherNext Cyclones Mini on the free v5e-1 TPU runtime. Inside, it auto-loads model weights from the storage bucket, loads initial-state weather data such as HRES initial conditions, initializes the WN2 (FGN) architecture, runs auto-regressive rollout steps to produce a forecast, and visualizes outputs like temperature, wind speed, and geopotential height. It can also run the direct tracker for cyclone tracks and compute training loss for a gradient step. Running the larger WN2 or WeatherNextCyclones checkpoints needs a v5p TPU or an H100 GPU, with the attention implementation switched as the notebook shows; only the Mini checkpoints run on a P100.
Strengths
- ✓Ships four operational-grade checkpoint families — WN2, three WeatherNext Cyclones generations, and two lightweight Minis — instead of just one model.
- ✓The README documents exactly which checkpoint ran live in production: WeatherNextCyclones_<2025, during the 2025 Atlantic hurricane season, publicly called FNV3.
- ✓Three separate access paths for raw forecast data — Google Cloud, WeatherLab, OpenMeteo — mean you don't have to run any code just to get daily output.
- ✓The Mini checkpoints are explicitly sized for a single TPU or GPU, so you can try the pipeline without H100-class hardware.
- ✓Code and Colab notebooks are Apache 2.0; the weights and other materials are CC BY 4.0 — both permissive licenses.
Important Considerations for WeatherNext 2
- △The README explicitly calls this research code "provided as-is," with no API stability guarantee — future updates can introduce breaking changes without notice, and the maintainers recommend pinning to a release like v0.3.0.
- △It isn't an officially supported Google product, per the repo's own disclaimer, and hasn't been produced with or endorsed by any government meteorological agency, so it can't substitute for official alerts or warnings.
- △Running anything beyond the Mini checkpoints needs real hardware — an H100 GPU or TPU v5p — and the README notes the implementation is optimized for TPU, so GPU users must switch the attention implementation themselves.
- △Full training needs the ERA5 dataset from ECMWF plus WeatherBench2's HRES data, both governed by separate terms and conditions you're responsible for checking compliance with.
The WeatherNext Family: GraphCast and GenCast
Frequently Asked Questions about WeatherNext 2
WeatherNext 2 has already run operationally: the README documents that its WeatherNextCyclones_<2025 checkpoint tracked live during the 2025 Atlantic hurricane season, publicly known as FNV3. Even so, Google's own disclaimer calls this experimental research code, not an officially supported product, and says it doesn't replace official meteorological agency alerts.
You can pull WeatherNext 2 output without running any code, through Google Cloud (Earth Engine, BigQuery, and Vertex AI), WeatherLab (which includes cyclone tracks), or OpenMeteo's API and interactive builder, per the README's own list of access points.
The README recommends TPU, since the implementation is optimized for it, and says the non-Mini WeatherNext 2 and WeatherNext Cyclones checkpoints need an H100 GPU for enough VRAM if you run on GPU instead. The lightweight Mini checkpoints are the exception — they're built to manage inference on a P100 or a single TPU.
The Colab notebooks and associated code are under Apache License 2.0, while everything else — weights, data, and other materials — is under Creative Commons Attribution 4.0 International (CC BY 4.0), per the repository's LICENSE section.
WeatherNext 2 forecasts cyclones using the same algorithm as the dedicated WeatherNext Cyclones models — the README notes the only functional difference is that WN2 also predicts 100m wind, since the two were trained independently.
Full WeatherNext 2 training needs the ERA5 reanalysis dataset from ECMWF, which the README recommends pulling as Zarr through WeatherBench2, plus WeatherBench2's HRES data for operational fine-tuning. Both datasets carry their own separate terms and conditions you're responsible for checking.
Who should try it — and who should skip
Reach for WeatherNext 2 if you want to run or study an operational-caliber AI cyclone and weather model with real deployment history — the WeatherNextCyclones_<2025 checkpoint tracked the 2025 Atlantic hurricane season live. Also reach for it if you just want forecast data and don't want to run anything: Google Cloud, WeatherLab, and OpenMeteo all serve it directly. Skip it if you need a stable, supported API — the README is explicit that this is unversioned research code — or if you lack TPU/H100-class hardware and only want the full-resolution checkpoints; the Mini models are the only ones sized for a single P100.
Curious whether weathernext is right for you?
Let ChatGPT, Claude, or Perplexity look into it — click below and see what AI actually says about weathernext.