We are excited to announce a key step in making our data-driven weather model even more accessible to a broader audience of developers, scientists, and weather enthusiasts. Already you can access real-time charts and data for the AIFS, and now the latest deterministic version of the AIFS (AIFS-single), v0.2.1, is available on Hugging Face so you can run the AIFS on your own device. Hugging Face is a widely used platform and community hub for sharing machine learning models. It provides an open ecosystem where developers can share, access, and experiment with state-of-the-art models across disciplines. By hosting the AIFS v0.2.1 model checkpoints on Hugging Face, we’re expanding the accessibility and transparency of our work at ECMWF, and we invite the community to join us in exploring this advanced AI-driven forecasting model.

Screenshot showing the AIFS model page on the Hugging Face platform.
Commitment to open science
We’ve been sharing our code and research (Lang et al. 2024), and now, by publishing AIFS model weights, we’re enhancing transparency, reproducibility, and collaboration, as part of our ongoing commitment to open science and data accessibility. We make this model available under a Creative Commons Attribution 4.0 International (CC BY 4.0) licence, meaning it can be used for any purpose. It allows researchers and practitioners worldwide to use, and experiment with, our model in new ways. We are really excited to see what people find!
Using AIFS weights in your forecasting workflow
The model weights are accompanied by a detailed ‘model card’ on Hugging Face, which includes details regarding the implementation and evaluation of the AIFS model. To demonstrate how to generate a forecast, the model card includes a link to a notebook which can be run on systems like Google Colab. This notebook offers step-by-step instructions and code snippets to guide you through the setup process.
Here’s a summary of the key steps to start using AIFS v0.2.1:
- Install the required packages
- Retrieve initial conditions from ECMWF Open data
- Generate forecasts using anemoi-inference
- Inspect the generated forecast
In the notebook, we demonstrate how to use the anemoi-inference NumPy-to-NumPy API for generating forecasts. However, anemoi-inference also provides other APIs, such as a command-line interface for customised workflows.
For more details about these APIs and their configurations, check the documentation.
Looking ahead
We’re excited and curious to see how the community leverages this powerful tool and to witness the contributions it inspires towards a more open, innovative, and data-driven future in weather forecasting.
Looking ahead, we’re also committed to expanding ECMWF’s Hugging Face model collection. Our goals include adding probabilistic models, updating models as new versions are released and providing complete checkpoints, which will provide optimiser states to easily fine-tune for a different application.