otter-xl-crps

An Otter weather forecasting model, from experiment rft_final_2026-05-18_20-44-majestic-aardvark_1e6 at step 18977, trained with code commit 6fd16b33cf0c02e6b9b61f5f5c9b0e8314adf12b.

• Grid: 121 x 240 (1.5 degree) ERA5.
• Forecasts 6-hourly steps autoregressively, from 4 6-hourly ERA5 states.
• Initial conditions: WeatherBench2 ERA5, gs://weatherbench2/datasets/era5/1959-2023_01_10-6h-240x121_equiangular_with_poles_conservative.zarr.

Usage

Forecasting needs a CUDA GPU.

git clone https://github.com/cambridge-mlg/otter && cd otter && uv sync
uv run python -m otter.hub forecast jonasscholz/otter-xl-crps \
    --init_time 2020-07-01T00 --output forecast.zarr

The forecast is a zarr store following the WeatherBench2 conventions. The config is instantiated with Hydra, which imports and calls the code it names: pin the revision with --revision <commit hash>.

Ensemble

This model was fine-tuned with a CRPS loss on MC dropout ensembles, so it forecasts with dropout on: each run of it samples an ensemble member. forecast runs 10 members by default, as in validation (--ensemble_members changes this, --seed the samples), and saves them along a realization dimension. Average them for the ensemble mean. With dropout off, the model drifts cold.

License

The weights are released under the MIT License, as is the code; see LICENSE.

Citation

@article{diaconu2026otter,
  title={Otter Weather: Skillful and Computationally Efficient Medium-Range Weather Forecasting},
  author={Diaconu, Cristiana and Scholz, Jonas and Shysheya, Aliaksandra and Markou, Stratis and Mukhopadhyay, Payel and Cranmer, Miles and Turner, Richard E},
  journal={arXiv preprint arXiv:2606.26421},
  year={2026}
}