amarchakitus/aurora-1.5-coreml

Aurora 1.5 for Apple's Neural Engine (Core ML)

Microsoft's Aurora 1.5 (0.25°, 1.26 B parameters; checkpoint aurora-0.25-v1.5.ckpt from microsoft/aurora), taken apart into 54 Core ML programs that run on the Neural Engine of an iPhone, iPad or Mac, as the AI Forecast Factory app runs it. The app downloads this repo and verifies every file against files.json.

File	What it is
encoder.mlmodelc, decoder.mlmodelc	The Perceiver encoder and decoder (fp16)
block_<layer>_<i>.mlmodelc	The 48 Swin blocks, each over a chunk of its windows (8-bit weights)
merge_*.mlmodelc, split_*.mlmodelc	The U-Net's patch merges and splits (fp16)
tables.safetensors	Per-patch constants (static-field and position embeddings), window layouts, the level-aggregation scores' float32 weights
static.safetensors	Land-sea mask and surface geopotential, 720 x 1440
aurora.json	Variables, normalisation, between-step clipping
manifest.json	Programs, stages and chunk sizes

Everything that does not depend on the weather is folded into constants: the 6-hour lead time's conditioning and every adaptive layer norm, the Perceivers' constant queries, the static fields' embedding. Over a 6-hour step against the IFS analysis, it scores within 0.4% of the float32 model on z500 RMSE.

Licence and attribution

Model and weights: Aurora © Microsoft Corporation, under the MIT licence (LICENSE). Changes made: the network rearranged into Core ML programs with constants folded in; the Swin blocks' weights quantised to 8 bits (per output channel); no retraining.

This is not a Microsoft product, and forecasts made with it are not Microsoft's.

Citation

Aurora 1.5 is fine-tuned from Aurora: Bodnar, C. et al. (2025). A foundation model for the Earth system. Nature. doi:10.1038/s41586-025-09005-y.