amarchakitus/aifs-single-v2-mlx

AIFS Single v2 for Apple devices (MLX + Core ML)

ECMWF's AIFS Single v2, converted to run on iPhone and iPad: the inference weights in half precision for MLX, the processor compiled for the Neural Engine with Core ML, the model graph, and the regridding operators the app uses on device. The app downloads this repo on first launch and verifies every file against files.json.

File	What it is
weights.safetensors	Inference weights, float16 (converted from the ECMWF checkpoint)
coreml/	The 16-layer processor as compiled Core ML models
graph.safetensors	Encoder/decoder graph and node attributes
grid.safetensors	The N320 grid's latitudes and longitudes
opendata.safetensors	ECMWF open data (0.25°) -> N320 interpolation operator
display.safetensors	N320 -> 0.25° interpolation operator, for maps
manifest.json	Bundle format, shapes and variable order

Licence and attribution

Model weights: AIFS Single v2 © European Centre for Medium-Range Weather Forecasts (ECMWF), licensed under CC BY 4.0. Changes made: converted from the PyTorch checkpoint to float16 safetensors and Core ML; no retraining.

Interpolation operators: built from ECMWF's earthkit-geo matrices (Apache License 2.0).

This is not an ECMWF product, and forecasts made with it are not ECMWF forecasts.

Citation

Lang, S. et al. (2024). AIFS – ECMWF's data-driven forecasting system. arXiv:2406.01465.