The Artificial Intelligence Forecasting System (AIFS) is a data-driven weather forecasting model developed by the European Centre for Medium-Range Weather Forecasts (ECMWF). It uses a graph neural network (GNN) encoder and decoder, along with a sliding window transformer processor, and is trained on ECMWF's ERA5 re-analysis and operational numerical weather prediction (NWP) analyses. AIFS is designed to be flexible and modular, supporting various levels of parallelism to enable training on high-resolution data. It produces highly skilled forecasts for upper-air variables, surface weather parameters, and tropical cyclone tracks. AIFS is run four times daily alongside ECMWF's physics-based NWP model and forecasts are available to the public under ECMWF's open data policy. AIFS is implemented in Python using PyTorch and PyTorch Geometric frameworks. It includes the Anemoi framework, a toolbox for data-driven weather forecast models. The model uses a combination of deep learning architectures, including vision transformers, neural operators, and graph neural networks. The first version of AIFS was introduced in October 2023, with an updated spatial resolution of approximately 0.25° in February 2024. The model uses a pre-norm transformer with shifted window attention and GELU activation function. AIFS is trained to produce 6-hour forecasts using ERA5 data and operational NWP analyses. It undergoes pre-training on ERA5 data from 1979 to 2020, followed by rollout training on ERA5 data from 1979 to 2018, and then fine-tuning on operational real-time IFS NWP analyses. The model uses a mixed precision training approach with a batch size of 16, and training takes about one week with 64 GPUs. AIFS produces accurate forecasts for various weather variables, including geopotential, temperature, wind speed, and precipitation, with improved performance compared to the ECMWF's Integrated Forecasting System (IFS). AIFS demonstrates highly competitive forecast performance for both upper-air and surface variables, with improved accuracy compared to IFS and ERA5. It produces accurate tropical cyclone track forecasts and shows better scores in the troposphere up to 100 hPa. However, it exhibits lower forecast accuracy at day 1 when verified against NWP analyses, but this degradation is not present when verified against radiosonde observations. AIFS also shows reduced forecast activity compared to IFS, with a smoothing effect on forecast fields with lead time. The current implementation of AIFS has certain limitations, including blurring of forecast fields at longer lead times due to the MSE training objective. Further improvements include refining the training loss or training probabilistic models to alleviate smoothing effects. AIFS is modular and scalable, with ongoing research focusing on extending it to probabilistic ensemble forecasting and data-driven regional modelling. The AIFS source code and weights will be releasedThe Artificial Intelligence Forecasting System (AIFS) is a data-driven weather forecasting model developed by the European Centre for Medium-Range Weather Forecasts (ECMWF). It uses a graph neural network (GNN) encoder and decoder, along with a sliding window transformer processor, and is trained on ECMWF's ERA5 re-analysis and operational numerical weather prediction (NWP) analyses. AIFS is designed to be flexible and modular, supporting various levels of parallelism to enable training on high-resolution data. It produces highly skilled forecasts for upper-air variables, surface weather parameters, and tropical cyclone tracks. AIFS is run four times daily alongside ECMWF's physics-based NWP model and forecasts are available to the public under ECMWF's open data policy. AIFS is implemented in Python using PyTorch and PyTorch Geometric frameworks. It includes the Anemoi framework, a toolbox for data-driven weather forecast models. The model uses a combination of deep learning architectures, including vision transformers, neural operators, and graph neural networks. The first version of AIFS was introduced in October 2023, with an updated spatial resolution of approximately 0.25° in February 2024. The model uses a pre-norm transformer with shifted window attention and GELU activation function. AIFS is trained to produce 6-hour forecasts using ERA5 data and operational NWP analyses. It undergoes pre-training on ERA5 data from 1979 to 2020, followed by rollout training on ERA5 data from 1979 to 2018, and then fine-tuning on operational real-time IFS NWP analyses. The model uses a mixed precision training approach with a batch size of 16, and training takes about one week with 64 GPUs. AIFS produces accurate forecasts for various weather variables, including geopotential, temperature, wind speed, and precipitation, with improved performance compared to the ECMWF's Integrated Forecasting System (IFS). AIFS demonstrates highly competitive forecast performance for both upper-air and surface variables, with improved accuracy compared to IFS and ERA5. It produces accurate tropical cyclone track forecasts and shows better scores in the troposphere up to 100 hPa. However, it exhibits lower forecast accuracy at day 1 when verified against NWP analyses, but this degradation is not present when verified against radiosonde observations. AIFS also shows reduced forecast activity compared to IFS, with a smoothing effect on forecast fields with lead time. The current implementation of AIFS has certain limitations, including blurring of forecast fields at longer lead times due to the MSE training objective. Further improvements include refining the training loss or training probabilistic models to alleviate smoothing effects. AIFS is modular and scalable, with ongoing research focusing on extending it to probabilistic ensemble forecasting and data-driven regional modelling. The AIFS source code and weights will be released