GemAI v3
The latest extended-range weather forecasting model from Salient. Operational and reforecast datasets now available.
16 gridded output variables cover key surface fields and atmospheric dynamics. GemAI v3 goes beyond standard snapshots with native sub-timestep aggregations like min/max temperature, wind gusts, and more. See documentation for details.
2m instantaneous, plus timestep min, max, and mean; 2m mean dewpoint.
Mean precipitation rate and snowfall rate.
10 m and 100 m mean wind speed; 10 m max gusts; 100 m u/v instantaneous components.
Surface downward shortwave flux; total cloud cover.
Mean sea-level pressure; 500 hPa geopotential.
ENSO, MJO, QBO, ZMZW, and more (available via API).
15-day benchmark CRPS scores against ECMWF ENS, ECMWF AIFS, and Google WeatherNext. Scores computed over 242 dates from July 2025 to March 2026, masked to land areas. GemAI v3 outperforms ECMWF ENS across all variables, is competitive with other state-of-the-art AI models on large-scale dynamics, and achieves best-in-class performance on 2m temperature, particularly the sub-timestep aggregations.
46-day benchmark CRPS scores against ECMWF ENS extended-range and a climatology baseline. Scores computed for 104 dates from July 2023 to July 2025, masked to land areas. GemAI v3 outperforms ECMWF ENS throughout the forecast horizon and retains positive or neutral scores relative to climatology at all leads.
pip install "xarray>=2025.6.0" "zarr>=3" "pcodec>=1" s3fsimport xarray as xr
ds = xr.open_zarr(
"https://gemv3-reforecast.salient-open-data.com/forecast",
chunks=None,
)
t2m = ds.sel(
forecast_date="2025-01-01",
lat=42.36, lon=-71.06, method="nearest",
)["2m_temperature"]
t2m = t2m.assign_coords(valid_time=t2m.forecast_date + t2m.lead
).swap_dims(lead="valid_time")
t2m.plot(hue="sample", add_legend=False)See docs for additional details.
Publicly accessible, for verification and backtesting.
Live forecasts from January 2026 onward.