FourCastNet 3
#5 in Wetenschop- & Bio-KICaltech · v3 · 5× · tolest 27. Aug. 2026
84
Momentum
FourCastNet 3 (FCN3) is a probabilistic AI weather forecasting model by NVIDIA, developed in collaboration with Caltech (Anima Anandkumar), Lawrence Berkeley National Laboratory, and UC Berkeley. It uses a purely convolutional neural network architecture designed for spherical geometry to generate global ensemble weather forecasts. The model delivers forecast accuracy that surpasses classical ensemble models and competes with diffusion-based approaches while running 8 to 60 times faster. It is freely available via Hugging Face, NVIDIA NGC/NIM, and as training code (Makani), and is licensed for both commercial and non-commercial use.
Momentum-Verloop
01.06.30.08.
Features
| Deployment Model | Self-hosted via NIM container (Docker, on-prem/cloud/workstation) or direct model download via Hugging Face/NGC for own infrastructure |
| Use Case Scope | Global weather forecasting (medium-range to subseasonal, up to 60 days), ensemble forecasting, climate modeling, early warning systems |
| Integrations | Earth2Studio (Python SDK), NVIDIA NIM microservices, Makani training framework, NGC Catalog, Hugging Face |
| License | Model weights: Apache 2.0 (Hugging Face); NIM container: NVIDIA AI Product Agreement / NVIDIA AI Foundation Models Community License |
| Platform | NVIDIA Earth-2 platform, runs on a single GPU (e.g. H100) or multi-GPU cluster (up to 1024+ GPUs for training) |
| Price | Free (open model, checkpoints and code freely available) |
| Release Date | July 16, 2025 (arXiv paper); NGC checkpoint v0.1.0 released July 29, 2025 |