Multi-Modal Foundation · nvidia
FourCastNet 3
A geometric deep learning model for rapid global probabilistic weather and subseasonal-to-seasonal forecasting.
What FourCastNet 3 is
FourCastNet 3 is a global weather forecasting model developed by NVIDIA that implements a geometric machine learning approach for probabilistic ensemble prediction. Designed to respect spherical geometry, the model addresses the challenge of accurately capturing spatially correlated uncertainty and preserving stable physical spectra across multiple atmospheric scales.
The system produces 60-day global forecasts at a 0.25-degree resolution in under four minutes on a single GPU. It delivers forecasting accuracy that surpasses leading conventional ensemble systems and rivals diffusion-based methods while operating 8 to 60 times faster, providing meteorologists and climate researchers with a tool for large-scale ensemble modeling and early warning predictions.
How it works
FourCastNet 3 uses a spherical neural operator architecture built as a fully convolutional neural network based on group convolutions defined directly on the sphere. By combining both local and global spherical convolutions, the architecture captures multiscale atmospheric dynamics while preserving energy spectra and numerical stability across extended forecast rollouts. Large-scale training across 1,024 GPUs and more is facilitated by a hybrid model- and data-parallel strategy inspired by domain decomposition methods from classical numerical weather prediction.
The model was trained on hourly estimates from the ERA5 reanalysis dataset covering the years 1980 to 2015, using a total dataset of 110,960 data points partitioned into 95% training, 2.5% validation (2018–2019), and 2.5% testing (2016–2017). It operates on a 0.25-degree equirectangular grid (721 x 1440 points) spanning 72 surface and pressure-level variables, including winds, temperature, geopotential height, specific humidity, mean sea-level pressure, and total column water vapour.
Key facts
- Architecture
- Spherical Neural Operator (fully convolutional group convolutions on the sphere)
- Parameters
- 710,867,670
- Input grid resolution
- 0.25° (721 x 1440)
- State variables
- 72 atmospheric surface and pressure-level variables
- Forecast temporal step
- 6-hourly resolution up to 60 days
- Training dataset
- ERA5 reanalysis (1980–2015)
- Testing dataset
- ERA5 reanalysis (2016–2017)
- Validation dataset
- ERA5 reanalysis (2018–2019)
- Inference speed
- 60-day forecast in under 4 minutes on a single GPU
- Model version
- v1
What you can use it for
- Medium-Range Ensemble Forecasting
- Generate rapid probabilistic ensemble weather predictions spanning days to weeks at 0.25-degree global resolution.
- Subseasonal-to-Seasonal Prediction
- Produce extended rollouts up to 60 days to study large-scale atmospheric phenomena and subseasonal climate trends.
- Meteorological Early Warning Systems
- Drive high-member ensemble pipelines rapidly to assess risks and uncertainties for extreme weather events.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Executes a complete 60-day global forecast at 0.25-degree resolution in under four minutes on a single GPU.
- Maintains realistic kinetic energy spectra and rollout stability over extended time horizons up to 60 days.
- Operates between 8 and 60 times faster than competing diffusion-based and conventional ensemble forecasting methods.
- Achieves high probabilistic calibration through a spherical convolutional architecture tailored to spherical geometry.
Watch out for
- Requires execution in a 64-bit Linux environment on compatible NVIDIA GPU architectures (Turing, Ampere, or Hopper).
- Requires a fixed 6D tensor input structure covering exactly 72 prescribed atmospheric variables on a 721 x 1440 grid.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use FourCastNet 3 commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
How fast does FourCastNet 3 run during inference?
On a single GPU, FourCastNet 3 computes a 60-day global trajectory at 6-hourly temporal steps and 0.25-degree spatial resolution in under four minutes.
What numerical precision is recommended for FourCastNet 3?
The model is designed to run using automatic mixed precision (AMP) with bfloat16 (bf16), while input and output tensors are typically formatted as float32.
What is the training dataset used for the model?
FourCastNet 3 was trained on the ECMWF ERA5 reanalysis dataset spanning 1980 through 2015, which provides atmospheric variables originally mapped on a 30 km grid resolving 137 atmospheric levels.
How to cite it
FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale, arXiv:2507.12144v2
Written from FourCastNet 3's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.