GraphCast operational for Apple's Neural Engine (Core ML)
Google DeepMind's GraphCast operational
(0.25°, 13 pressure levels; checkpoint GraphCast_operational - ERA5-HRES 1979-2021 - resolution 0.25 - pressure levels 13 - mesh 2to6 - precipitation output only), taken apart
into six 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 |
|---|---|
grid.mlmodelc |
Grid node embedding and the encoder edges' sender projection |
enc_edges.mlmodelc |
Encoder (grid to mesh) edge MLP |
mesh_embed.mlmodelc |
Multimesh edge embedding |
mesh_edges.mlmodelc, mesh_nodes.mlmodelc |
Processor edge and node updates; weights as inputs, one program for all 16 layers |
decoder.mlmodelc |
Decoder (mesh to grid) and output head |
weights.safetensors |
The processor's weights, fp16 |
tables.safetensors |
Mesh node embedding; per-edge low-rank coefficients of the encoder's and decoder's constant edge terms |
graph.safetensors |
The three graphs' indices and the multimesh edges' features |
static.safetensors |
ERA5 land-sea mask and surface geopotential, 721 x 1440 |
graphcast.json |
Grid, input and output channels, normalisation |
What does not depend on the weather is folded into constants: the mesh nodes' embedding, each edge set's embedding (kept per edge as 8 or 16 coefficients on a basis, to 1e-5), the edge MLPs' sender and receiver projections done per node. Gains are folded into the weights for fp16 on the Neural Engine. Over a 6-hour step it matches the float32 model to 0.4-0.6% of each field's change, closer than the bfloat16 model Google runs.
Licence and attribution
Model weights: GraphCast © Google DeepMind (Google LLC), under the Creative Commons Attribution 4.0 International licence (CC BY 4.0), per Google's notice of 6 August 2026 (which replaces the CC BY-NC-SA 4.0 text still in the checkpoint's metadata). Changes made: the network rearranged into Core ML programs, constants folded and compressed (low-rank), weights converted to fp16 and scaled for fp16 arithmetic; no retraining.
GraphCast was trained on ECMWF's ERA5 and HRES data: modified Copernicus Climate Change Service information. Neither the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus information or data it contains. The static fields here are ERA5's.
This is not a Google product, and forecasts made with it are not Google's.
Citation
Lam, R. et al. (2023). Learning skillful medium-range global weather forecasting. Science 382, 1416-1421. doi:10.1126/science.adi2336.