NVIDIA Earth-2 has unveiled FourCastNet3 (FCN3), a groundbreaking AI-powered global weather forecasting system that delivers exceptional accuracy and speed for large ensemble predictions. The model represents a major advancement in data-driven weather modeling, offering enhanced probabilistic performance, computational efficiency, and spectral fidelity across subseasonal timescales. FCN3 achieves a 60-day forecast with 0.25° spatial resolution and 6-hourly updates in under four minutes on a single NVIDIA H100 Tensor Core GPU. This marks an 8x improvement over GenCast and a 60x speedup compared to traditional numerical weather prediction systems like IFS-ENS. The system’s ability to maintain realistic atmospheric patterns over extended lead times sets it apart, with ensemble members preserving physical consistency and spectral characteristics even at 60 days. The model’s architecture leverages a fully convolutional, spherical neural operator framework. Unlike its predecessor, FourCastNet2, which relied on spherical Fourier neural operators, FCN3 integrates both local and spectral convolutions. These operations are parameterized using Morlet wavelets, enabling localized filters for atmospheric phenomena while ensuring computational efficiency through NVIDIA CUDA optimizations. This design allows FCN3 to capture complex spatial relationships and maintain stability over long forecasts. A key innovation in FCN3 is its stochastic approach, which introduces variability through a hidden Markov model. By conditioning on a spherical noise variable, the system generates ensemble forecasts efficiently at each step, avoiding the computational overhead of diffusion-based methods. It is trained using a composite loss function that combines spatial and spectral metrics, ensuring accurate learning of atmospheric dynamics and uncertainty patterns. To address scalability, FCN3 employs a novel model-parallelism strategy inspired by domain decomposition techniques in traditional weather modeling. This allows the model to train on up to 1,024 GPUs simultaneously, distributing spatial operations via the NVIDIA Collective Communications Library (NCCL). The approach reduces memory bottlenecks and enables larger, more complex models to be processed effectively. In terms of performance, FCN3 outperforms physics-based systems like IFS-ENS and matches the predictive accuracy of GenCast. Its probabilistic ensembles exhibit spread-skill ratios near one, reflecting well-calibrated forecasts where uncertainty estimates align with real-world variability. Diagnostics such as rank histograms confirm that ensemble members remain consistent with observations, validating their reliability. The model’s spectral fidelity is particularly notable. It accurately reproduces atmospheric energy cascades and sharp weather patterns, even at 60-day lead times. For example, FCN3’s predictions of wind intensities during Storm Dennis in 2020 retained the correct spatial and frequency characteristics, avoiding the high-frequency blurring or noise artifacts common in other ML models. FCN3’s training and inference tools are accessible through NVIDIA NGC, with a pre-trained checkpoint available for immediate use. Users can run predictions via Earth2Studio, a platform designed for AI weather modeling. For custom applications, the code is hosted in the Makani repository, with recommendations to install torch-harmonics and use automatic mixed-precision (bf16) for optimal performance. Developed by a team including researchers from NVIDIA, Lawrence Berkeley National Laboratory, Caltech, and other institutions, FCN3 aims to bridge the gap between traditional numerical models and data-driven AI systems. By combining speed, accuracy, and physical realism, it offers a scalable solution for subseasonal forecasting, which is critical for applications like disaster preparedness and climate studies. The system’s ability to handle large ensembles while maintaining computational efficiency positions it as a significant tool in the evolving AI-driven weather prediction landscape.