Successor to Aardvark (ICLR 2026) that advances the skill–compute Pareto frontier: the deterministic model beats the best NWP baseline by 9.6% at 24h trained on under 3.5 A100-days, while Otter-XL beats IFS ENS by 9.7% CRPS and GenCast by over 2% at roughly 10x less compute. Transfers out-of-the-box to an acoustic-scattering PDE task. Fully Cambridge author list (Turner group).

Model Details

Architecture DENSE

Paper

Venue ICLR 2026
scienceresearch

Related